a robot application for marine vessel inspectiondownload.xuebalib.com/xuebalib.com.36861.pdf ·...

A Robot Application for Marine Vessel Inspection

bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull bull

Markus EichDFKI - Robotics Innovation Center 28359 Bremen Germanye-mail markuseichdfkideFrancisco Bonnin-Pascual Emilio Garcia-Fidalgo and Alberto OrtizUniversity of the Balearic Islands 07122 Palma de Mallorca Spaine-mail albertoortizuibesGabriele BruzzoneISSIA - CNR (Via De Marini 6) 16149 Genoa Italye-mail gabrielebruzzonegeissiacnritYannis KoveosGlafcos Marine Ltd 18537 Piraeus Greecee-mail ykovaiosglafcos-marinecomFrank KirchnerDFKI - Robotics Innovation Center 28359 Bremen Germanye-mail frankkirchnerdfkide

Received 8 March 2013 accepted 18 November 2013

Seagoing vessels have to undergo regular inspections which are currently performed manually by ship sur-veyors The main cost factor in a ship inspection is to provide access to the different areas of the ship since thesurveyor has to be close to the inspected parts usually within armrsquos reach either to perform a visual analysisor to take thickness measurements The access to the structural elements in cargo holds eg bulkheads isnormally provided by staging or by ldquocherry-pickingrdquo cranes To make ship inspections safer and more cost-efficient we have introduced new inspection methods tools and systems which have been evaluated in fieldtrials particularly focusing on cargo holds More precisely two magnetic climbing robots and a micro-aerial vehiclewhich are able to assist the surveyor during the inspection are introduced Since localization of inspection datais mandatory for the surveyor we also introduce an external localization system that has been verified in fieldtrials using a climbing inspection robot Furthermore the inspection data collected by the robotic systems areorganized and handled by a spatial content management system that enables us to compare the inspection data ofone survey with those from another as well as to document the ship inspection when the robot team is usedImage-based defect detection is addressed by proposing an integrated solution for detecting corrosion and cracksThe systemsrsquo performance is reported as well as conclusions on their usability all in accordance with the outputof field trials performed onboard two different vessels under real inspection conditions Ccopy 2014 Wiley PeriodicalsInc

1 INTRODUCTION

For obvious reasons large tonnage vessels such as bulkcarriers dry cargo ships or tankers (see Figure 1) undergoregular inspections to prevent structural damage thatcan compromise the vesselrsquos integrity These inspectionsare usually performed in accordance with an inspectionprogram that depends on the requirements of the so-calledclassification societies (in short the classes) and they arecomprised of visual close-up surveys as well as thicknessmeasurements obtained by means of nondestructive testingmethods (NDTs) (Tanneberger amp Grasso 2011) For a close-up survey the surveyor has to get within armrsquos reach ofthe part under observation for adequate visual inspectionStructural damage pitting and corrosion are visually

estimated based on the experience of the surveyor and theinspection process is usually documented using cameras totake images chalk and pen for defect marking and a clip-board for note taking Some solutions based on unmannedunderwater vehicles (UUVs) have been proposed recentlyfor the inspection of underwater areas eg the hybrid ROVsolution by ECA Robotics (httpwwweca-roboticscom)the HAUV by Bluefin (Kaess Johannsson Englot Hoveramp Leonard 2010) and the VideoRay ROVs (httpwwwvideoraycom) The first two systems are primarilyintended for an underwater hull survey while the latter isfor water tank inspection

Regarding the inspection of dry areas providing accessto the relevant parts of the ship eg inside the cargo hold of

Journal of Field Robotics 31(2) 319ndash341 (2014) Ccopy 2014 Wiley Periodicals IncView this article online at wileyonlinelibrarycom bull DOI 101002rob21498

320 bull Journal of Field Roboticsmdash2014

Figure 1 Illustration of traditional inspection methods (left) general cargo ship (center) staging previous to inspection and(right) cherry-picking [Source (leftcenter) Lloydrsquos Register (right) httpwwwstandard-clubcom]

a bulk carrier is the most time-consuming part of the inspec-tion process As can be seen in Figure 1 (center) traditionalship surveying methods include prior to the survey theinstallation of scaffolding to allow the surveyor to inspectstructures such as bulkheads beams stiffeners and brack-ets which are usually several meters above the bottom of thehold Besides the scaffolding ldquocherry-pickingrdquo methods arealso employed in this case the surveyor reaches the point ofinterest inside a basket transported by a tower crane or bya hydraulic arm [Figure 1 (right)] Clearly these procedurescan be dangerous for the surveyor For this reason and be-cause of the high costs of gaining access to a ship for inspec-tion the EU-funded research project MINOAS (Caccia et al2010b) set up a consortium to introduce robots into the shipsurveying process The basis of the consortiumrsquos expertisecomprised two classification societies different marine ser-vice providers and partners involved in robotics researchThe key idea was to introduce a set of novel tools to enhancethe ship surveying process For the interested reader a moredetailed discussion of the application scenario can be foundin Ortiz et al (2010)

This paper reports the results of the introduction ofheterogeneous robots to the area of close-up surveys of thestructural elements of large-tonnage vessels where most ofthe work is still performed manually The effort is a mix-ture of novelty and integration and the proportion of eachis different for every platform Nevertheless the main con-tribution is the fully integrated inspection system cover-ing all the stages of an inspection procedure based on theuse of robots and supporting software something that didnot exist before the project MINOAS We introduce a het-erogeneous robot team with different locomotion abilitiesnamely a micro-aerial vehicle (MAV) a lightweight mag-netic crawler supported by an external positioning unitand a heavyweight magnetic crawler equipped with a ma-nipulator arm for thickness measurement Two additionalsystems were developed to assist the surveyor a spatial con-tent management system (SCMS) to host and present in acentralized way the inspection data collected by all the plat-forms and a visual defect detection solution for automatic

defect detection They are all described in the followingsections

2 REENGINEERED INSPECTION PROCEDURE

On the basis of the expertise of the classes involved andthe maritime industry and as part of the MINOAS projectworking plan a total of three stages were defined in or-der to implement an inspection procedure based on robotsand compatible with the requirements of the classes (Tan-neberger amp Grasso 2011)

Stage 1 Fast visual inspection overview The goal of thisstage is to cover large areas of selected parts of the inner hullsupplying visual inspection data to get a rough overviewabout the state of the vessel and searching for defects suchas coating breakdown corrosion and cracks The imagescollected must be tagged with pose information since theareas of interest for Stages 2 and 3 are visually selectedin this phase Due to the aforementioned requirement anaerial platform turns out to be the best option In more detailthe vehicle must be able to perform vertical stationary andlow-speed flight in indoor environments

Stage 2 Visual close-up survey The procedure in Stage 2is to get a better impression of the coating and the possibledamages and mark the defective areas in order to repairthe damage or perform thickness measurements The vehi-cle must be capable of moving on horizontal sloped andvertical ferromagnetic surfaces At this stage the camera ofthe robot has to be steady and very close to the structuralparts in order to provide high-resolution high-quality im-ages Proper lighting of the area under inspection shouldbe provided Marking the defects directly on the struc-ture once they are confirmed by the surveyor would be anoptional feature of the robot The acquired images haveto be tagged with positioning information Finally ease ofsystem setup is also required Under these constraints alightweight magnetic crawler was considered to be a suit-able platform

Stage 3 Thickness measurement collection At selectedparts identified during Stage 2 the thickness of the material

Journal of Field Robotics DOI 101002rob

Eich et al A Robot Application for Marine Vessel Inspection bull 321

Figure 2 (Left) bulk carrier hold (right) shape and size of a shell frame

Table I Relation between systems and inspection stages

System Stage 1 Stage 2 Stage 3

Aerial platform timesLightweight inspection system

(crawler amp localization device)times

Heavyweight inspection system(crawler amp NDT thicknessmeasurement device)

times

Defect detection system times timesSpatial contents management system times times times

is determined by taking measurements at selected points ofthe hull The locomotion speed is not important and thesetup requirements are permitted to be more onerous thanfor Stages 1 and 2 The main goal of Stage 3 is to collect thick-ness measurements at selected points of the hull structuresMotion over horizontal sloped and vertical ferromagneticsurfaces is still required Compared to Stage 2 a higher pay-load capability is required in order to transport thicknessmeasurement devices and a suitable manipulator arm Theclimbing robot has to be able to move inside shell frames ofthe type shown in Figure 2 Constraints on the vehicle sizeare imposed by the need for maneuvering between the T-shaped shell frames of the size indicated in Figure 2 (right)A heavyweight inspection system was selected to fulfill therequirements for this stage Table I specifies the relation be-tween the different systems and the mission stage wherethey are used

The MINOAS inspection procedure places suitablerobots into the inspection process but is generic enough tocover all the requirements of a traditional close-up surveyand becomes therefore the basis for a fully robot-based

inspection procedure for marine vessels Neverthelesssince a very significant fraction of the inspection effort at thestructural level (as specified by the rules set for the classes)is expended on cargo holds and on the different elementsthey involve eg bulkheads stiffeners and cross-decks thedevelopments and tests in this paper refer mostly to cargoholds

3 MINOAS INSPECTION PLATFORMS

31 Related Work on Inspection Platforms

This section reviews previous work that could match therequirements for the MINOAS platforms Due to their dif-ferent natures aerial platforms and crawling systems areconsidered separately the latter referring jointly to both thelightweight and the heavyweight inspection systems

311 Aerial Platforms

MAVs have been increasingly popular as robotic platformsin recent years Their development has been driven by com-mercial research government and military purposes Thiskind of vehicle enables access to hazardous environmentsthat are usually difficult to reach by humans or groundvehicles These robots are an adequate solution for inspec-tion tasks in eg remote or safety-compromised areas Forthese platforms to achieve autonomy a full navigation so-lution is required Recently a number of navigation solu-tions have been proposed for multirotors including plat-form stabilization self-localization mapping and obstacleavoidance They differ mainly in the sensors used to solvethese tasks the amount of processing that is performedonboardoffboard and the assumptions made about theenvironment The laser scanner has been used extensivelydue to its accuracy and speed For instance DryanovskiValenti amp Xiao (2013) and Grzonka Grisetti amp Burgard



(2012) proposed full navigation systems using laser scanmatching and IMU fusion for motion estimation embed-ded within SLAM frameworks that enable such MAVs tooperate indoors In Bachrach Prentice He amp Roy (2011)and Dryanovski Valenti amp Xiao (2013) a multilevel ap-proach is described for three-dimensional (3D) mappingtasks Infrared or ultrasound sensors are other possibili-ties for implementing navigation solutions Although theytypically have less accuracy and require higher noise tol-erance several researchers (Bouabdallah Murrieri amp Sieg-wart 2005 Matsue Hirosue Tokutake Sunada amp Ohkura2005 Roberts Stirling Zufferey amp Floreano 2007) haveused them to perform navigation tasks in indoor environ-ments since they are a cheaper option than laser scan-ners Vision-based navigation has become quite popularfor MAVs lately The success of cameras in general roboticscomes mainly from the richness of the sensor data suppliedcombined with their low weight low power designs andrelatively low prices Nevertheless for the particular caseof MAVs the associated higher computational cost has ledresearchers to find optimized solutions that can run overlow-power processors Among the most recent papers pub-lished in this regard some propose visual SLAM solutionsbased on feature tracking either adopting a frontal mono orstereo camera configuration [eg Fraundorfer et al (2012)]or choosing a ground-looking orientation [eg ChowdharyJohnson Magree Wu amp Shein (2013)] Others focus on ef-ficient implementations of optical flow calculations eitherdense or sparse and mostly from ground-looking cameras[eg Zingg Scaramuzza Weiss amp Siegwart (2010)] or theydevelop methods for landing tracking and taking off us-ing passive [eg Meier et al (2012)] or active markers [egWenzel Masselli amp Zell (2011)] also using a ground-looking camera

312 Climbing Robot Platforms

The field of wall-climbing robots which naturally turns outto be relevant for this application has received a certainamount of attention since the late 1990s (Silva amp Tenreiro2010) Referring specifically to marine applications a roboticsystem that was developed to inspect hot welding seamswas introduced by Shang Bridge Sattar Mondal amp Bren-ner (2008) This small-sized system has a weight of 30 kgrequires a safety rope during operation and uses an infraredsensor to check the temperature of the hot seam after weld-ing Heavyweight hull cleaning robots have also been usedfor ship surveying and repair These robots weigh more than100 kg and are used to remove large areas of coating on theshiprsquos hull using water jetting techniques or brushes (Ortizet al 2007) Some robots are already available for marine in-spection services such as the robot Steel-Climber from MikoMarine (Miko 2013) or the Magnet Crawler M250 fromJetstream Europe (Jetstream 2013) Both are heavyweightmagnetic crawlers for blasting ship cleaning and inspec-

tion The robot CROMSKI (Jung Schmidt amp Berns 2010) isused for dam inspection and is able to climb vertical wallsindependently of the material of the wall Another robot us-ing a tracked system with suction pads is described in KimSeo Lee amp Kim (2010) The robot is a self-contained systemthat also integrates the motor that produces the vacuumAnother suction-pad-based climbing approach is describedin Raut Patwardhan amp Henry (2010) It enables the robotto walk on glossy and flat surfaces for window cleaning inbuildings and it also relies on clean surfaces An exampleof a robot using magnetic wheels for inspection purpose isdescribed in Tache et al (2009) The adaptability of the sys-tem is provided by different joints that allow the adjustmentof the robotrsquos kinematics A tracked robot using permanentmagnets is described in Kalra Guf amp Meng (2006)

32 Aerial Inspection Robot for Stage 1

321 General Overview and Design of the Robot

The MAV prototype is based on the well-known Pelicanquadrotor from Ascending Technologies [see Figure 3(a)]This is a 50-cm-diameter platform with 254 cm propellersable to carry a payload of 650 g and equipped with a stan-dard navigation sensor suite a barometric pressure sensorfor height estimation a GPS receiver and a full three-axisinertial measuring unit (IMU) Furthermore the MAV has aHokuyo lightweight laser scanner with a range of up to 30m which is used not only for obstacle detection but alsoby deflection of lateral beams using mirrors to estimate thedistance to the floor as well as to the ceiling Visual informa-tion is collected by means of a flexible vision system withan appropriate structure for supporting one ground-lookingcamera and two additional units which can be tailored forthe particular inspection mission to be performed such astwo forward-facing cameras forming a stereo vision systemone camera facing forward and the other facing upward orto save weight a single camera facing forward Apart fromthe onboard controllers the vehicle carries an additionalhigh level processor (HLP) that obviates the need to sendsensor data to a base station since it processes them on-board thus avoiding any communications latency insidethe critical control loops

The configuration shown in Figure 3(a) includes a Core-Express board fitted with an Intel Atom 16 GHz processorand 1 GB RAM The different sensors are attached to theHLP through a USB Finally communications are imple-mented through a WiFi link The wireless device attachedto the vehicle is connected to the HLP using a dedicatedPCI Express port avoiding the need for wireless communi-cations to share USB bandwidth

322 Control Architecture

As on similar platforms the control software architecturecomprises at least two physically separated agents the MAV



Figure 3 The aerial vehicle (a) platform overview (b) self-localization and mapping modules (c) mission execution modules (d)example of a mission specification file

itself and a ground station More specifically the differentcomputational resources of the MAV run the control algo-rithms as follows (either as firmware or as software) (i)as is well known in the Pelican the main ARM7 low-levelcontroller (LLC) runs the low-level software taking careof attitude stabilization and direct motor control (Gurdanet al 2007) (ii) the secondary ARM7 high-level controller(HLC) runs the position controller described in AchtelikAchtelik Weiss amp Siegwart (2011) and (iii) the HLP exe-cutes on top of the robot operating system (ROS) runningover Linux Ubuntu ROS nodes providing platform motionestimates as well as platform safety interaction with the on-board platform controllers and WiFi communication withthe ground station Finally in our configuration the groundstation comprises a cluster of laptops running ROSLinuxUbuntu to perform offboard operations

Figures 3(b) and 3(c) depict the control software run-ning on the HLP and on the ground station It features

self-localization mapping obstacle avoidance and path-planning modules as well as mission control and super-vision modules Self-localization as a central capabilityfor this platform is implemented following a 2D laser-based motion estimation approach (in accordance with therequirements of Stage 1) In more detail the control soft-ware permits operation in both semiautonomous and au-tonomous modes In the first mode of operation an oper-ator is expected to send velocity commands in x y andz using the sticks of a remote control (RC) unit while thevehicle provides hovering and height control functional-ities using the onboard sensors and the low-level atti-tudeposition controllers In the second mode of operationthe vehicle performs autonomously missions described bymeans of mission specification files [MSFs see Figure 3(d)for a very simple example] In short MSFs are parsed inorder to identify and perform the actions requested (go-tonavigate-to and take-photo) making use of the sensor data



Figure 4 The lightweight crawler (a) isometric view with 1) a video camera and LED light source 2) an elastic tail joint 3) amagnetic rear wheel rocker 4) a high-power LED for position tracking 5) a magnetic front wheel 6) a 12 V LiPo battery pack and7) a 24 GHz M-PCM receiver (b) vehicle realization

processing components laser preprocessing vertical posi-tion and odometry whose results feed the SLAM and thenavigation components together with the low-level con-trollers Finally a safety manager implements a number ofsafety behaviors such as monitoring go-to and navigate-to ac-tions before sending the corresponding motion commandsto the HLC preventing the robot from flying too high or tooclose to the ceiling and monitoring the battery voltage

Among the different components enumerated abovethe ground station runs those control modules that can toler-ate latency in the communications namely simultaneous lo-calization and mapping path planning and mission execu-tionsupervision while critical control loops run onboardthe vehicle in order to ensure minimum delay a require-ment also reported by other authors (Achtelik BachrachHe Prentice amp Roy 2009) to permit autonomous flyingFurthermore information exchange between the laptops isperformed by wire and the wireless datalink is left onlyfor communication with the MAV That is to say only onelaptop talks directly with the MAV which reduces multi-ple point-to-point communications for the same data butthey are republished by this laptop to provide the full clus-ter with all the information This configuration allows newcomputers to be added to the cluster as needed ensuringthere is no extra wireless communication with the vehi-cle For more details about the MAV control software referto Bonnin-Pascual Garcia-Fidalgo amp Ortiz (2012)

33 Lightweight Inspection Robot for Stage 2


The lightweight climbing robot is depicted in Figure 4 It isactuated by two 12 V DC motors that drive the two frontwheels on which a total of 112 neodymium magnets areattached The adhesion force on an iron surface is 1216N per magnet The polarities of the magnets are orientedalternately to increase the adhesion force Each wheel con-

sists of two rows of magnets with a foam material appliedin between to increase the traction during climbing Themagnets are integrated into flexible rubber holdings thatprovide adaptability to the surface and allow the system totraverse between surfaces Furthermore a tail is attached tothe system using a flexible tail joint During wall climbingthe tail bends around the z-axis allowing motion in circleswith a small radius At the tail tip a rocker has been incor-porated with two additional magnetic rings This providesan additional degree of freedom that allows the tail wheelsto adapt to uneven parts and climb on corrugated metalparts Finally since the system weight is well below 1 kgno safety precautions need to be taken except for a smallcatching net to be used by a second person to catch the robotin case it drops from eg a bulkhead This is in contrast tomany existing magnetic climbing robots which need to besecured by a crane or a safety rope As a result the systemcan be quickly deployed to obtain visual close-up data fromareas that are hard to reach

To provide visual data a camera is attached to the frontof the robot including an LED light source to enhance theimage quality The robot also stores high-resolution imagesand video streams directly on a local SD-card for postpro-cessing The lightweight crawler is also capable of mark-ing defects directly on the bulkhead of a ship using amicropump to apply acrylic varnish on the surface SeeEich amp Vogele (2011) for more details on the design of thelightweight crawler


This section describes the control software of the lightweightinspection system a graphical overview of which is shownin Figure 5 As can be observed the diagram also in-cludes the SCMS although the corresponding details willnot be given until Section 5 since it is a system that is or-thogonal to the robot team We include the SCMS in this



Figure 5 Overview of the control software of the lightweight inspection system In addition to the operator as man-in-the-loopthe system architecture is comprised of the mobile robot and the 3D localization unit The drawing also gives the relation betweenthe lightweight crawler and the SCMS

Figure 6 User interfaces for the FPV of the lightweight inspection system (left) small and portable hand-held 5-cm screen device(center) custom developed screen box comprising a 58 GHz video receiver and an 18-cm TFT screen (right) full immersive videogoggles used by the surveyor during field trials

discussion to illustrate how in particular the lightweightcrawler interacts with it Any other platform interacts in ap-proximately the same way thanks to the modularity prin-ciples that have been adopted for the design of the full in-spection system1 With regard to the lightweight inspectionsystem in particular the operator controls the robot using a24 GHz M-PCM RC unit which provides the speed signalto the motors and gets direct feedback from the robot viaa first-person-view (FPV) interface The remote control alsoprovides a PCM interface to the control computer allowingdifferent control modalities eg direct control by an opera-tor where the robot receiver sends the commands to the DCdrivers or indirect control by a navigation algorithm Thepump spray unit is actuated using a microactuator which is

1The same would apply to the optical tracking unit and the heavy-weight crawler

also directly triggered via the RC The FPV interface is im-plemented through an analog wireless 58 GHz link Dur-ing field trials two different options were evaluated (seeFigure 6) hand-held devices and video goggles the lat-ter providing the operator with an immersive view of thescenery through the camera of the robot Because maneuver-ing the robot only from the FPV can cause nausea especiallyif the robot is at a significant height it has been observedthat it is desirable for a direct line of sight to exist betweenthe operator and the robot

333 Associated Vision-based Localization Unit

Due to the size and weight of the lightweight crawler whichwere directly imposed by the requirements the systemcannot carry additional sensors that could be used to de-rive its position However in Stage 2 metric localization is



Figure 7 3D localization unit (left) integrated system (right) control architecture

mandatory in order to localize the acquired data The 3Dlocalization system and the control architecture are shownin Figure 7 In this diagram U corresponds to the camerarsquoshorizontal image resolution and V to its vertical one Thecamera is in the fixed reference frame of the laser rangefinder and both are actuated in accordance with the esti-mated pitch and yaw angles (θ ϕ) of the target by the twoservo motors of the pan-tilt unit (PTU) The image providedby the camera is preprocessed using a Laplacian of Gaus-sian filter (LoG) (Haralick amp Shapiro 1992) The resultingimage g(u v) has steep intensity gradients and is well suitedto detect the bright light source of the LED in the imagesThe output of the detector is the position (u v) of the cen-ter of gravity (CoG) of the image Common problems whentracking a light source are that the position might jumpto other bright sources during the tracking process or thetarget itself can get temporarily occluded

Due to its well-known robustness to multimodal dis-tributions and occlusions (Isard amp Blake 1998) we haveadopted a Monte Carlo particle filter that is used in com-bination with sequential importance resampling (SIR) Themethod we use for the blob estimation in the image planeis a variant of the sample-based Monte Carlo localization(MCL) method which is well-known for robot pose track-ing within a 2D map (Thrun Fox Burgard amp Dellaert 2000)In general the following recursive equation is solved usingthe particle filter in order to estimate the position of a system(μ is a normalization factor)

Bel(xt ) = μ1

σradic

2πeminus 1

2 ( ot minusxt σ )2

timesint

p[xt |xtminus1T (xtminus1 xtminus2) g(xtminus1 σ )]Bel(xtminus1)dxtminus1

(1)

In contrast to the application of the particle filter in robotlocalization no direct action update a is available to thesystem therefore we estimated the last action by the trans-lation between the two positions To account for the changeof speed and direction between two pose estimation stepswe add Gaussian noise into the motion model The last mo-

tion at time xtminus1 is thus based on the translation betweenthe two prior estimated positions T (xtminus1 xtminus2) The Gaus-sian noise is incorporated as the function g(xtminus1 σ ) In ourcase we take the distance between the estimated positionand the new blob measurement given by the blob detectionalgorithm

In the sequential importance resampling step eachsample is redrawn from the sample set according to itsweight ωi = μ 1

σradic

2πexp[minusot minus xi

t 2(2σ 2)] which ensuresthat good samples are reproduced while samples with alow likelihood are removed from the set2 The output ofthe system is thus the 3D position of the robot in polarcoordinates which is projected onto the reference frameof the tracker using x = d sin θ cos ϕ y = d sin θ sin ϕ andz = d cos θ This estimate is regularly provided to the oper-ator and to the SCMS

34 Heavyweight Inspection Robot for Stage 3


As mentioned before unlike the lightweight climber thiscrawler requires a larger payload capacity in order to carryan electrical robotic arm and an ultrasound probe for non-destructive thickness measurement apart from other sen-sors and devices As a consequence the mechanical designof this heavier vehicle has been essentially determined bythe choice of the climbing system To minimize the risk ofdetaching and crashing while using a passive adherencesystem magnetic tracks were preferred to wheels by virtueof their ability to maximize the contact area with the metalsurface The result was the Magnetic Autonomous RoboticCrawler (MARC) shown in Figure 8

As can be seen a cart actuated by a DC brushless ser-vomotor and rigidly connected to a couple of wheels hasbeen mounted at the rear of the vehicle The wheeled cartpositioned in home up and down configurations as shownin Figure 8 helps the vehicle to cross slope discontinuitiesand to not overturn when working on vertical slopes As far

2μ ensures the weights ωi sum up to 1



(a) Wheels in home position used for opera-tions over non-vertical slopes

(b) Wheels in up position used for handlingslope discontinuities between 45 and 90 de-grees

(c) Wheels in down position used whenworking on vertical slopes

Figure 8 MARC rear wheel configurations

as the magnetic tracks are concerned custom-made platelet-shaped magnets have been designed built and directlyfixed to the holes of the track chain brackets Each magnetwhich measures 40 times 22 mm exerts a maximum magneticforce of about 100 N Since under nominal operating con-ditions eight to ten magnets per track are in contact withthe metallic surface the corresponding attractive force is of800ndash1000 N per track (1600ndash2000 N in total) Each track isactuated by a 140 W brushless motor coupled to a gearheadwith a reduction ratio 851 for a resulting maximum speedof 012 ms Motors and gear boxes are mounted at the bot-tom part of the vehicle together with Li-Ion battery packsplaced along the longitudinal axis and at the rear part of theplatform


The MARC sensor system consists of the following a high-performance miniature attitude heading reference system(AHRS) with GPS combining MEMS sensor technology anda highly sensitive embedded GPS receiver providing ori-entation inertial and when possible GPS measurementsand four laser range sensors able to measure the distanceto eg lateral bulkheads and allowing the computation ofthe platformrsquos position and orientation inside an operatinglane The vehicle-embedded real-time computing platformis based on commercial-off-the-shelf (COTS) hardware andfree software in particular a single-board computer hard-ware architecture possessing PC-compatible CPUs hasbeen adopted to support a Linux-operated system config-ured for real-time performance as discussed in BruzzoneCaccia Bertone amp Ravera (2009)

The MARC control system is able to follow a linearsurface on the basis of lateral range measurements Themeasurements provided by the pair of laser range sensorsmounted perpendicularly to the vehicle side allow us toestimate the distance d and orientation ϕ of the vehicle withrespect to a linear vertical shell frame The proposed control

law for following a linear surface using a simple Lyapunov-based design is

u = ulowastϕ = minuskdu

lowast(d minus dlowast) minus kϕ sin ϕ(2)

where ulowast is the linear reference speed of the vehicle ϕ is itsyaw rate dlowast is the reference range from the tracked linearsurface d is the current distance to the surface and kd andkϕ are positive constants This allows MARC to follow alinear surface at a reference range remaining parallel tothe tracked surface See Bibuli et al (2012) for more detailsabout MARC

343 Associated Robotic Arm for NDT Measurement

Given the complexity of ship structures and their commonenvironmental conditions eg highly corroded rusted oreven bent structural members ultrasonic thickness mea-surement (UTM) becomes a tedious task even for skilledUT operators involving a number of phases surface prepa-ration which can become the longest task since it is notuncommon to have significant scale and rust on the metal-lic surface application of couplant usually a high viscosityliquid such as glycerin whose purpose is to increase thetransmission coefficient of the medium traversed by theultrasonic pulse that goes from the probe to the uneven sur-face of the specimen and UT probe application and mea-surement which can require additional surface preparationif the UTM is not steady or if unreliable readings are ob-tained Due to these conditions and in order to requireonly a minimum of skills from the user the robotic armcarried by MARC exhibits a certain degree of adaptabilityand intelligence in order to allow the measurement of awide range of significant ship structural components suchas frame stiffeners brackets and face plates as requiredby the classes Five degrees of freedom (four angular andone linear) a holding torque of 1 kgm (185V) and anangular resolution of 03 are the key resulting specifications



of the arm at the mechanical level Regarding the arm con-troller it features a USB interface with MARC and an ARMCortex M3 (72 MHz) processor running a 400 Hz controlloop that in particular implements spline-based trajectoryfollowing Finally regarding the end-effector it comprisesa grinder for surface preparation (using current monitoringfor torque control) a peristaltic pump to inject the couplantthat improves the transmission of the ultrasonic pulse andthe corresponding ultrasound sensor Referring particularlyto the task of thickness measurement the robot is equippedwith an FPGA-based board responsible for the high voltagepulse generation (lt 390 V) the return signal amplification(0ndash100 dB) and the digitization at a maximum samplingrate of 100 MHz and 10-bit resolution For a more detaileddescription refer to Koveos Y Kolyvas T amp Drikos (2012)UTM relies on measuring the time for an ultrasonic pulseto propagate along the material being tested and extract itsthickness on the basis of the speed of sound c It is typicalfrom NDT to use A-Scan waveforms ie ultrasonic ampli-tude versus time of arrival for visual inspection Howeversince the raw waveform is usually noisy and may containsignals from multiple reflections time is measured froma filtered signal yn representing the energy of the ultra-sound This is obtained by low-pass filtering a rectified inputsignal xn

Since the waveform is buffered a zero-phase formula-tion is used to preserve phase information (in the followingP and Q are respectively the feedforward and the feedbackfilter orders bi and ai are the corresponding filter coeffi-cients and N is the number of samples)

1) Forward filtering

ylowastn = 1

a0

⎛⎝ Psum

i=0

bi |xnminusi | minusQsum

j=1

ajy1nminusj

⎞⎠ n = 0 N

2) Reverse-time filtering

yn = 1a0

⎛⎝ Psum

i=0

biylowastNminusnminusi minus

Qsumj=1

ajynminusj

⎞⎠ n = 0 N

3) Time reversingyn = yNminusn n = 0 N

(3)Regarding thickness calculation it can be implementedthrough a peak detection algorithm leveraging the fact thatultrasonic echoes are monotonically decreasing due to ab-sorption and reflection losses time of flight can be estimatedas the time between those peaks Despite the fact that thismethod is simple it relies on single points of the waveformand is subject to variations of the user-selected parame-ters eg filter bandwidth amplification etc A second al-gorithm based on the autocorrelation of the unfiltered inputsignal is preferred instead Although more computationallyintensive it provides robust estimations due to the use ofthe entire waveform minimizing any dependency on theuser-provided parameters and augmenting the level of au-

tonomy The thickness is then estimated as follows

thickness = c

2dt

(arg max

j

sumn

xnxnminusj

) j isin nmin nmax

(4)with dt the sampling period and nmin and nmax definedby the minimum and maximum measurable thicknesseswhich in this work come from a wide range 3ndash30 mmsatisfying the surveyrsquos requirements typically 5ndash25 mm

4 A VISION-BASED SOLUTION FOR DEFECTDETECTION

This section proposes a vision-based integrated solutionfor detecting two kind of defects of metallic surfaces thatare generally considered relevant to determining the con-dition of a vessel namely coating breakdowncorrosionand cracks Bonnin-Pascual (2010) performed a thoroughreview of the different techniques that have been used sofar for defect detection in general Some of those propos-als are specifically related to vessel inspection althoughthey mostly refer to the inspection of the external part ofthe hull by means of (mainly tethered) unmanned under-water vehicles Their main goal is to assist with the detec-tion of the loss of the external coating aquatic life adheringto the hull (to prevent future corrosion) artificial objectsattached to the hull (to avoid sabotage) weld inspectionetc

The approach described in this section aims at detect-ing the aforementioned defects from the digital images cap-tured by the MINOAS robotic platforms It must be regardedas an assistance tool in the hands of the surveyor which es-sentially means there will always be a human making thelast decision about the state of the corresponding structuralcomponent More specifically the defect detector is an inte-grated solution that considers both kinds of defect throughan algorithm comprising two stages The first stage corro-sion detection is in charge of labeling all areas of the inputimage suspected of being affected by corrosion The secondstage crack detection makes use of the pixels classified ascorroded as clues for detecting cracks The rationale behindthis approach comes from the observation that typicallyin metallic surfaces cracks are surrounded by corroded ar-eas so that guiding the crack detector in accordance withthe output of the corrosion detector becomes a highly effec-tive strategy for enhancing its performance The two stagesare outlined in the following sections For more detailssee Bonnin-Pascual (2010)

41 Corrosion Detection

The corrosion detector is based on a supervised classifica-tion scheme that comprises two steps that can be consid-ered as two weak classifiers The idea is to chain differentfast classifiers with poor performance in order to obtaina global classifier attaining a higher global performance



The first classifier is based on the premise that a corrodedarea exhibits a rough texture Roughness is measured asthe energy of the symmetric gray-level co-occurrence ma-trix (GLCM) calculated for down-sampled intensity valuesbetween 0 and 31 for a given direction α and distance d The texture energy E is defined as E = sum31

i=0

sum31j=0 p(i j )2

where p(i j ) is the probability of the occurrence of graylevels i and j at distance d and orientations α or α + π Patches with an energy lower than a given threshold τE iethose that exhibit a rough texture are finally candidates tobe more deeply inspected

The second classifier operates over the pixels of thepatches that have survived the roughness step This classi-fier makes use of the color information that can be observedfrom the corroded areas unlike the first classifier It worksover the hue-saturation-value (HSV) space after the realiza-tion that HSV values that can be found in corroded areasare confined to a bounded subspace of the HS plane Al-though the V component has been observed to be neithersignificant nor necessary to describe the color of the corro-sion it is used to prevent the well-known instabilities in thecomputation of hue and saturation when the color is closeto white or black This second step requires a prior trainingstep to learn the color of the corroded areas which consistsin building a bidimensional histogram of HS values for im-age pixels labeled as ldquocorrodedrdquo The resulting histogram issubsequently filtered by zeroing entries whose value is be-low 10 of the highest peak and next applying a bilateralfilter considering the binsrsquo heights as the intensity valuesof a digital image This approach filters the histogram us-ing a kernel consisting of two Gaussians one for the spatialdomain and another for the range domain The color clas-sifier proceeds as follows for every 3-tuple (h s v) giventhresholds mV MV and mS pixels close to black v lt mV or white v gt MV and s lt mS are labeled as noncorrodedwhile the remaining pixels are classified as corroded if theHS histogram contains a nonzero value at the (h s) entry

42 Crack Detection

The crack detector proceeds in accordance with a perco-lation model similarly to the approach described in Yam-aguchi amp Hashimoto (2010) although their detector wasconstructed for searching cracks in concrete this fact makesthe authors assume a geometrical structure that does notmatch exactly the shape of the cracks that are formed in steelThe percolation model used is developed from a region-growing procedure that starts from a seed pixel and prop-agates in accordance to a set of rules In our case the rulesare defined to identify dark narrow and elongated sets ofconnected pixels which are then labeled as cracks Once aseed pixel has been defined the percolation process startsas a two-step procedure during the first step the percola-tion is applied inside a window of N times N pixels until thewindow boundary is reached in the second step if the elon-

gation of the grown region is above εN a second percola-tion is performed until either the boundary of a window ofM times M pixels (M gt N ) is reached or the propagation cannotprogress because the gray levels of all the pixels next to thecurrent boundary are above a threshold T Finally all thepixels within the region grown are classified as crack pixelsif the region elongation is larger than εM

Within the N times N - or the M times M-pixel window thepercolation proceeds in accordance to the next propagationrules (i) all eight neighbors of the percolated area are de-fined as candidates and (ii) each candidate p is visited andincluded in the percolated area only if its gray level valueI (p) is lower than a threshold T which has been initializedto the seed pixel gray level At the end of the percolationprocess the mean gray level of the set of pixels is checkedto determine if the region is dark enough to be considered acrack otherwise the region is discarded and another perco-lation starts at a different seed pixel Seed pixels are definedover a regular grid with a step of pixels and are requiredto coincide with an edge whose gray level is below γs Toensure that the relevant edges are always considered a di-lation step follows the edge detection where the dilationthickness is in accordance with Furthermore since thecrack detector operates under the guidance of the corro-sion detector as mentioned above seed pixels are requiredto lie within corroded areas The different conditions forbeing a seed pixel taken together speed up the crack detec-tion process and reduce the false positive rate thanks to thecorrosion-based guidance

5 SPATIAL CONTENT MANAGEMENT SYSTEM FORROBOT-ACQUIRED INSPECTION DATA

Classification societies use software packages such as theGermanischer LLoyd HullManager3 which provides themeans to assess a shiprsquos condition and store the relevantdata for later use Mobile tools for damage assessment alsoexist from Sertica4 With this tool the surveyor can incor-porate inspection data using a portable device Publicationsregarding inspection data management tools can be foundin Fletcher amp Kattan (2009) All these tools however requirean operator who manually inputs the inspection data

The SCMS that is introduced in this section integratesthe surveyors and the inspection robots in a sort of man-in-the-loop strategy The inspection data acquired by therobots are collected with the SCMS and can be evaluatedduring or after the inspection Currently no system of thiskind is available The SCMS is a central element of the in-spection system that is proposed in this work In more de-tail it permits collecting sharing and displaying in a cen-tralized manner all the information gathered using a 3D

3httpwwwgl-maritime-softwarecomgl-hullmanagerphp4httpwwwserticadk



ship

cargo

port

hold1

star-board port

hold2

star-board

bridge

super-structure

crane1

Figure 9 The Spatial Content Management System (left) the 3D user interface (right) the data are internally assigned to topologicalparts of the vessel Within the nodes the localization of the data is performed metrically

representation of the vessel under inspection All robot sys-tems introduced in this paper can provide images videosor thickness measurements during the inspection to theSCMS using common interfaces As mentioned in CacciaBibuli amp Bruzzone (2010a) a surveyor does not use a met-ric representation for localizing inspection data but ratheruses the spatial topology of the ship ie ldquocargo hold 1port side bulkhead left siderdquo inspection data are topolog-ically assigned to the named parts of the ship within theSCMS The ship as a whole is thus internally representedas a tree structure each node corresponding to a certainpart of the ship The 3D interface of the SCMS is shown inFigure 5 The data are internally assigned to the topolog-ical parts of the vessel (cf Figure 5) The inspection dataare anchored to the nodes of the tree structure This repre-sentation as a graph apart from being logical to the usersallows quick access to the data since all the inspection datafor a certain topological unit eg a certain cargo holdare located under a certain branch of the tree structureNevertheless the location of the inspection data is alsorepresented in a metrical way within the data node leafseach leaf node having its own local metric reference framealigned to the 3D representation of the ship The origin ofthe reference frame corresponds to a fixed position egwhere the localization unit was placed during the inspec-tion with the only constraint being that across surveysapproximately the same position must be used Within theSCMS the surveyor navigates through a 3D model of theship to get access to the different areas while the inspec-tion data items are projected metrically into the 3D modelas data blobs which can also be selected by the surveyorFurthermore all the data associated to a certain node ir-respective of their modality and recording date are avail-able once the item is selected Finally 3D submodels canbe associated to each tree node in order to obtain a higherlevel of detail of the corresponding structural element whenselected

6 SYSTEM PERFORMANCE EVALUATION

The marine inspection system described in this work wastested both in the laboratory and in field trials This sectionreports the results collected for every platform during thedifferent experiments that were performed The first trialwas performed on a bulk carrier ship that was under repairin a shipyard in Varna Bulgaria Several ship surveyorsattended the trials and provided comments on the usabilityof the various platforms and on possible improvements Thesecond trial was also performed on a container cargo shipin the same shipyard which served also as a final validationpoint for all systems The two test cargo holds are shown inFigure 10

61 Performance of the Aerial Platform

This section reports on the results of a number of labora-tory and field experiments to demonstrate the navigationcapabilities of the MAV These experiments were plannedto cover the kind of capabilities required to perform Stage 1of a MINOAS inspection mission To begin with Figure 11shows results for the MAV operating in semiautonomousmode with the operator trying to reach with the assistanceof the platform a set of waypoints within a laboratory ofthe University of the Balearic Islands [see Figures 11(a) and11(c)] and during the first field trial onboard the bulk car-rier ship [see Figures 11(b) and 11(d)] As can be noticedthe platform effectively provides the operator with heightcontrol and hovering functionalities The first experimentin autonomous mode that is reported whose results areprovided in Figures 12(a)ndash12(d) shows the ability of theplatform to reach without the intervention of an operator anumber of waypoints at different heights (within a buildingof the University of the Balearic Islands) More precisely themission describes a wall sweeping task as would be doneonboard a ship although at the scale of a laboratory Themission consists in attaining with an accuracy of 01 m a



Figure 10 Vessel holds involved in the first and the second field trials (left) cargo hold of the bulk carrier after it was cleanedand (right) cargo hold of the container ship

Figure 11 The MAV operating in semiautonomous mode (ab) snapshots of the environments (c) path followed by the MAVduring the laboratory experiment (d) path followed by the MAV during the field trial (The red square indicates the takeoff positionand the green circle the landing position)

total of ten waypoints along a zigzag path as indicated inFigure 12(b) A different kind of experiment (also performedwithin a building of the University of the Balearic Islands) isshown in Figures 12(e)ndash12(h) In this case apart from reach-ing a set of waypoints autonomously the goal is to per-form two short sweeping tasks in adjacent rooms so thata safe path through the door communicating booth rooms

has to be planned and followed during the flight That is tosay in this experiment the MAV faces all the difficulties ofa real scenario As in the previous experiment the differ-ent waypoints are correctly attainedmdashwithin the accuracyrequirements determined by the mission specification file(01 m)mdashwhile the local and global path planners producesafe paths toward the targets



Figure 12 MAV results in autonomous mode under laboratory conditions [first experiment] (a) snapshot of the environment (b)sequence of waypoints to be reached 0ndash910 (c) path followed by the MAV and (d) 2D path followed by the MAV superimposedover part of the map [second experiment] (e) snapshots of the two rooms where the experiment takes placemdashthe MAV departsfrom above goes below and finally returns abovemdash (f) sequence of waypoints to be reached 0ndash910 and (g) and (h) map and2D3D path followed by the MAV



Figure 13 MAV results in autonomous mode from the second field trial (a) hold area where the experiment takes place (b)sequence of waypoints to be reached by the MAV 0ndash657ndash158510 (c) and (d) map and 2D3D paths followed by the MAVand (e) pictures taken at waypoints 1 4 8 and 9 [from top to bottom]

A third set of experiments in autonomous modetaking place onboard the container ship during the secondfield trial is reported here [see Figure 13(a)] This timethe mission description file specified a sweeping taskconsisting in reaching a collection of 16 waypoints alonga zigzag-like path [see Figure 13(b)] and the experimentwas performed several times to compare the results of theconsecutive executions The results which can be found inFigures 13(c) and 13(d) show that (i) the paths estimatedby the vehicle are consistent with the environment as wellas the map built by the laser-based SLAM componentand (ii) the behavior of the MAV is repeatable (up tothe accuracy requested in the mission specification file01 m) both of which suggest the validity of the navigationapproach Pictures taken by the onboard cameras at someof the waypoints can be found in Figure 13(e)

62 Performance of the Lightweight InspectionSystem

621 Lightweight Crawler

As for the climbing capability of the robot several testswere performed over mainly vertical bulkheads and alsoalong 70 slopes Figure 14(a) shows the climbing robot on

a slope and Figure 14(b) shows it on a transverse bulk-head In this regard Table II shows the maximum tear-offforce against the maximum vertical drag force exerted bythe robot while climbing vertically on a bulkhead alonga straight line The reference measurement for the robotwas taken while climbing a clean bare steel frame with-out any coating The system was able to create a lift of198 N while climbing vertically The wheels of the robotblocked because the motors were not able to create moretorque Therefore this number corresponds to the max-imum lift the robot could produce under optimal con-ditions The maximum tear-off force was measured as451 N In all the experiments except for the referencemeasurement the wheels started to slip and the magnetscould not produce enough adhesive force for the robot tomove upward A live image of the crawler was received bythe small hand-held device depicted in Figure 6 (left) Incontrast to the laboratory experiments the 24 GHz analogvideo signal suffered from several problems being signif-icantly distorted at a distance of more than three metersBecause of the video transmission issue only the videosand images stored directly on the robot could be used bythe surveyor Of particular relevance in this regard is thefact that the robot was able to take high-quality snapshots



Figure 14 First trial for the lightweight crawler the climbing performance was tested on a 70 slope (a) as well as on a verticaltransverse bulkhead (b) (c)ndash(e) sample images taken by the robot

Table II Performance of the climbing robot on different ver-tical surfaces

MaxVertical Max

Drag Tear-off RobotSurface Condition Force (N) Force (N) Behavior

5 mm bare steelplate

clean 451 198 block

1 mm sheet metal clean 255 124 slipcoated steel clean 303 157 slipcoated steel corrosion 286 122 slipbare steel corrosion 255 171 slip

of some damage at some welding seams that were of interestto the surveyors because the damage was not visible to thesurveyor without using the robot [cf Figures 14(c)ndash14(e)]The quality of the images was sufficient to estimate the levelof corrosion The experimental setup during the trial withina container hold which involved a surveyor an operatorthe 3D tracking unit and the lightweight crawler is shownin Figure 15(a) The robot was constantly sending images tothe SCMS which was running on a laptop Some snapshotsof the transmitted images are shown in Figures 15(c) and15(d) To check the performance of the crawler an operatorwithin the line of sight controlled the robot while the sur-veyor would indicate where to move The robot was ableto climb on the vertical walls of the cargo hold In additionit was useful for inspecting the spaces behind the guidancerails for the containers [see Figure 15(b)] that were not visi-

ble to the surveyor without using the robot However rustand dirt would stick to the magnets causing problems forthe crawler [see Figure 15(e)]

Summarizing the trials we can say that on the onehand the main strength of the lightweight crawler is itsease of use which was proved by the surveyors them-selves being able to control the robot without prior train-ing Even without the tracking system and the SCMS therobot alone proved able to serve as an extension of the sur-veyorrsquos eyes supplying him with a quick look at parts ofthe ship that otherwise are accessible only using stagingor cherry-picking methods On the other hand the sec-ond trial also showed clearly the limits of using a mag-netic climbing system in a real environment rust and dirtdid not allow the crawler to climb freely on the bulkheadalthough it was possible for the operator to navigate therobot around such critical patches Since rusty patches pro-vide less adhesion force for magnets a system with strongermagnets would solve the problem such as the heavyweightcrawler MARC described in this paper The heavyweightcrawler had no problems in heavily corroded bulkheadsbut it required more safety precautions and more setuptime

622 Localization Experiments

In this section we report the performance of the 3D trackingunit while tracking the lightweight crawler To this end theabsolute error was measured and compared with a refer-ence value of 03 m This value comes from the structure ofcargo holds in different sorts of vessels (eg bulk carriers



Figure 15 Second trial of the lightweight crawler (a) experimental setup (b) view of the wall climbed by the robot (c) and (d)sample of transmitted video images from the inspection run and (e) state of the magnets covered in dirt and rust after one of theruns

11

12

13

14

15

16

17

18

19

3 4 5 6 7 8 9 10 11 12 13 14

Pos

ition

Y (

m)

Position X (m)

measurement error

x y z dist σ3866 minus3267 minus0771 5120 00264637 minus2039 minus0743 5120 00185013 minus0687 minus0774 5119 00295003 +0732 minus0797 5118 00354974 +0710 +1303 5190 00324996 minus0690 +1214 5187 00294619 minus2034 +1196 5187 00353844 minus3251 +1258 5189 0023

Figure 16 3D tracker performance (left) absolute positioning error as a function of distance (right) center of mass (x y z) andstandard deviation (σ ) for each cluster of the repeatability test (units are in meters)

and containerships) and corresponds to the average dis-tance between the sections in which bulkheads and wallsare typically divided In this way the surveyor can knowthe section from which the inspection data were taken Dur-ing the experiments the crawler would move along a ver-tical bulkhead of the cargo hold along a constant heightof 15 m Results from a first experiment can be found inFigure 16(left) As can be observed the resulting accuracy

is around 15 cm up to a distance of 11 m while when thetarget is more distant from the tracker the error of the sys-tem is significantly larger This is due to the resolution ofthe camera and the focal length of the optics To increase theoverall accuracy a different camera with higher resolutionand a different focal length could be used In this particu-lar case however the height of the cargo hold during thesecond field trial was around 12 m which indicates that



56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

minus5 minus4 minus3 minus2 minus1 0minus1

0

1

2

3

4

Y (m)

Z (

m)

Figure 17 Two trajectories estimated by the 3D tracker (first third) 3D perspective (second fourth) YZ projection

0 1000 2000 3000 4000 5000 600005

1

15MARC vertical bar following

r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]

Figure 18 (Left) MARC climbing a bulkhead inside the ship cargo hold (right) MARC telemetry while following a side framemeasured ranges and estimated distance distance error and orientation error (from top to bottom) where in the upper plot r1(green) refers to the front right sensor r3 (red) refers to the rear right sensor and d (black) is the estimated distance to the sideframe

the tracking performance was good enough to localize therobot everywhere inside the hold ie in accordance withthe requirements

A second experiment was carried out to evaluate therepeatability of the tracker Eight different positions wereselected on a bulkhead and the tracker was placed at adistance of around 5 m The target was placed randomlyat the selected positions and a total of 70 measurementswere taken by means of the tracker As expected the trackerproduced eight point clusters one for each distinct positionThe center of mass and the standard deviation for eachcluster are shown in Figure 16(right) Taking into accountthose results it becomes obvious that the tracking is highlyrepeatable Figure 17 shows the trajectories estimated by thetracker along two of the many runs that were performedduring the trials The area covered by the runs was around

20 m2 The tracking unit was located at a distance of around6 m from the bulkhead

63 Performance of the Heavyweight InspectionSystem

631 Heavyweight Crawler

The main goal during the ship trials was to demonstratethe ability inside a ship hold of the MARC (i) to climbvertical walls in general and bulkheads in particular and(ii) to follow vertical shell frames in autonomous modeThe ability of MARC to climb a vertical painted metal-lic wall inside a ship hold in a dirty environment in thepresence of rust was verified in a first set of experimentsDuring these preliminary tests the vehicle was positionedon the vertical walls with the support of human operators



Figure 19 (Left) MARC on a vertical bulkhead taking a thickness measurement during a field trial (right) screen shot of the UTsoftware

Figure 18(left) shows MARC climbing a bulkhead The ropevisible in the pictures is just a safety measure ie it wouldnot support the vehicle under normal working conditionsThe experiments performed consisted in MARCrsquos followinga vertical stiffener frame tracked with two of the lateral laserrange finders with which MARC is fitted while climbing abulkhead By way of illustration Figure 18(right) shows thevehicle telemetry while executing one of these experimentsThe different plots show from top to bottom the measuredranges and estimated distances and the range and orienta-tion errors As expected the vehicle manages to reduce thetracking error to zero after deliberately forcing the robot tomove closer to one of the sides

632 Thickness Measurements

To test MARCrsquos ability to measure thicknesses when fit-ted with the arm and the NDT end-effector the crawlerwas made to climb a vertical bulkhead [Figure 19(left)]while aligning itself to the stiffener frames using its laterallaser distance sensors Once in place the operator measur-ing the thickness with the ultrasound can use the roboticarm to perform the measurement The measuring proce-dure starts then by setting a rough estimate of the desiredpoint (x y z θ ) where the measure has to be taken

Next the arm starts to move along the previously de-fined vector until contact with the surface is detected bymeans of reed switches or the motorrsquos loading measureprovided by the servos which makes unnecessary any pre-vious knowledge of the exact relative location of the surfacepoint Since the angle θ is provided as a rough estimate of theorientation of the surface to be measured the actual valuemust then be identified This is essential for a reliable mea-surement since grinding and the UT probe placement arehighly sensitive to orientation misalignment reed switches

and the servorsquos angular sensors are again used to thisend

The next step starts grinding at a predefined load leveland time duration which must be selected by the operatordepending on the shiprsquos state and the amount of rust TheUT probe is next placed on the ground spot and a couplantis applied through the peristaltic pump suitable for highlyviscous fluids Finally the measurement is performed anda filtered A-Scan along the real UT waveform is presentedto the operator The measured waveform and its locationcan be stored prior to the start of a new measurementFigure 19(right) shows the results for one of the measure-ments where the UT waveform is in red and the A-Scan isin blue Numerous settings familiar to the UT operator suchas gain filter bandwidth time window etc can be alteredat any point in time although these settings are typically ad-justed during calibration at the beginning of the survey asrequired by the surveyor The figure shows for comparisonthickness estimates from both the peak detection algorithmsand algorithms based on the autocorrelation function Themeasurement 91 mm was verified by the shipyardrsquos rep-resentative Given the distant operation (typically above10 m height) the number of steps to perform and the ac-curacy of the operations required by the UTM the armrsquosmicro-controller software was designed and developed sothat the measuring process was as automatic as possiblerequiring minimal user intervention or knowledge of theenvironment

64 Performance of the Defect-detection Solution

The performance of the defect detector has been evalu-ated using digital images provided by the MINOAS part-ners or else collected from real vessels during field tri-als In general the different images were captured withoutspecific settings for the imaging process particularly with



(a)

Original image τE = 02 τE = 01 τE = 005

(b1) (b2) exec time = 17 ms (b3) exec time = 83 ms (b4) exec time = 224 ms

(c1) (c2) exec time = 21 ms (c3) exec time = 53 ms (c4) exec time = 206 ms

Figure 20 (a) Corroded areas detected using different energy threshold values τE (b) and (c) Image samples and correspondingdefect detections (1) original image (2) corrosion detector output (3) crack detector output and (4) crack detector output if thecorrosion stage is skipped

regard to the illumination Performance figures have beenobtained for each stage of the classifier using manually la-beled images as the ground truth Illustrative results of thedefect detector are provided in Figure 20 On the one handFigure 20(a) provides results for the same input image usingdifferent energy thresholds As can be observed τE can betuned to decrease false positives and just allow the detec-tion of the most significant corroded areas In the imagespixels labeled as corroded are color-coded as red orangegreen and blue in accordance to the likelihood provided bythe histogram with red for high likelihood and blue for lowlikelihood On the other hand Figures 20(b) and 20(c) showsthe results for the full defect detector The figure also com-pares the output of the crack detection stage with the outputof an unguided configuration As can be seen the number offalse positive detections is conspicuously reduced when thecrack inspection makes use of the output of the corrosiondetection

Quantitative performance results for the full set of im-ages have been obtained by calculating the percentages offalse positives FPno pixels and of false negatives FNnopixels On the one hand for the corrosion detector the re-sults are 980 (FP) and 586 (FN) while for the crackdetector the results are 072 (FP) and 057 (FN) On theother hand when the corrosion stage output is used to guide

the crack detector the result is a reduction in the FP from229 to 072 and a speed-up in the execution time of 31xThis performance is in accordance with our assumption thatmost of the cracks in a metallic surface appear in areas thatare also affected by corrosion and prove that using corro-sion to guide the crack detection results in a faster and morerobust detector Referring to the execution times the defectdetector provides corrosion-labeled images in 7ndash25 ms andcorrosion-guided crack-labeled images in 30ndash180 ms Theseexecution times correspond to images ranging from 120000to 172000 pixels and for a runtime environment including alaptop fitted with an Intel Core2 Duo processor (220GHz4GB RAM)

65 Evaluation of the Spatial ContentManagement System

The content management system was introduced duringthe second trial on the container ship The tracking systemwas placed in a fixed position within the cargo hold The 3Dmodel was assigned to the topological node of the vessel asdescribed in Section 5 and the position of the 3D tracker wascalibrated by measuring several of the distances from thebulkheads The lightweight crawler was used during the ex-periments Transmitted inspection data were synchronized



Figure 21 (a) Cargo hold where the SCMS evaluation took place Illustration of SCMS operation (b) 3D view (c) image at thedata node and (d) tree view

with the position provided by the tracking unit thus pro-viding information that was tagged with time and positionLive inspection views of the SCMS can be found in Fig-ure 21 For a start Figure 21(a) shows a 3D view of the holdthe position of the inspection robot and the data blobs thathad been recorded by the system Within this view the op-erator can directly select inspection data just by clicking onthe corresponding blob [Figure 21(b)] The SCMS can alsoshow a tree view [Figure 21(c)] where the operator can se-lect inspection data based on the recording time the type ofdata or search for specific data contents

7 CONCLUSION AND FUTURE RESEARCH

This paper has introduced a vessel inspection system com-prising a number of robotic systems fitted with differentlocomotion capabilities (aerial and magnetic crawling) andsensing modalities Auxiliary systems such as an integrateddefect detection solution and a spatial content managementsystem (SCMS) have been verified in field trials The abil-ities of each platform have been derived directly from therequirements imposed by a reengineered inspection proce-dure using robots consisting of three distinct inspectionstages An integrated defect detection solution has been

applied in order to select the critical spots for the use ofnondestructive testing methods during the third stage Theareas of improvement include apart from the industrial-ization of the platforms for a more robust operation theenhancement of the locomotion capabilities together withthe augmentation of the platformrsquos intelligence to increaseits level of autonomy The SCMS can also be enhanced bymeans of the semantic annotation of the structural parts inorder to lead to a more understandable representation ofthe inspection data eg ldquoblistering on the upper stool be-tween the third and fourth shell framerdquo as well as with theincorporation of 3D reconstruction functionality using eglaser-generated point clouds in order to avoid any depen-dence on the availability of CAD models of the vessel underinspection Finally the collaboration between the platformsis a topic for further research

ACKNOWLEDGMENTS

This work was funded by the European Commission un-der project MINOAS (SCP8-GA-2009-233715) FranciscoBonnin-Pascual and Emilio Garcia-Fidalgo have been par-tially supported by the European Social Fund through re-spectively grants FPI10-43175042V and FPI11-4312 3621R



(Conselleria drsquoEducacio Cultura i Universitats Govern deles Illes Balears) The authors want to thank Felix Grim-minger and Kristin Fondahl for their work on the hard-ware design of the lightweight crawler and the 3D trackingunit Finally the authors thank Giorgio Bruzzone EdoardoSpirandelli and Mauro Giacopelli for their extraordinarycontribution to the design development and testing of theMARC platform

REFERENCES

Achtelik M Achtelik M Weiss S amp Siegwart R (2011) On-board IMU and monocular vision based control for MAVsin unknown in- and outdoor environments In IEEE Inter-national Conference on Robotics and Automation (ICRA)(pp 3056ndash3063) Shanghai China

Achtelik M Bachrach A He R Prentice S amp Roy N (2009)Autonomous navigation and exploration of a quadrotorhelicopter in GPS-denied indoor environments In Interna-tional Aerial Robotics Competition (IARC) (pp 582ndash586)Mayaguez Puerto Rico

Bachrach A Prentice S He R amp Roy N (2011)RANGE-Robust autonomous navigation in GPS-deniedenvironments Journal of Field Robotics 28(5) 644ndash666

Bibuli M Bruzzone G Bruzzone G Caccia M GiacopelliM Petitti A amp Spirandelli E (2012) MARC Magneticautonomous robotic crawler development and exploita-tion in the MINOAS Project In International Conferenceon Computers and IT Applications in the Maritime Indus-tries (COMPIT) (pp 62ndash75) Liege Belgium

Bonnin-Pascual F (2010) Detection of cracks and corrosionfor automated vessels visual inspection Masterrsquos thesisUniversity of Balearic Islands Available at httpsrvuibesref1233html

Bonnin-Pascual F Garcia-Fidalgo E amp Ortiz A (2012) Semi-autonomous visual inspection of vessels assisted by anunmanned micro aerial vehicle In IEEERSJ InternationalConference on Intelligent Robots and Systems (IROS) (pp3955ndash3961) Vilamoura Algarve Portugal

Bouabdallah S Murrieri P amp Siegwart R (2005) Towardsautonomous indoor micro VTOL Autonomous Robots18 171ndash183

Bruzzone G Caccia M Bertone A amp Ravera G (2009) Stan-dard Linux for embedded real-time robotics and manufac-turing control systems Robotics and Computer-IntegratedManufacturing 25(1) 178ndash190

Caccia M Bibuli M amp Bruzzone G (2010a) Definition of nav-igational patterns Technical report CNR ISSIA GenoaItaly

Caccia M Robino R Bateman W Eich M Ortiz ADrikos L Todorova A Gaviotis I Spadoni F amp Apos-tolopoulou V (2010b) MINOAS A marine inspectionrobotic assistant System requirements and design In Sym-posium on Intelligent Autonomous Vehicles (IAV) (pp479ndash484) Lecce Italy

Chowdhary G Johnson E Magree D Wu A amp Shein A(2013) GPS-denied indoor and outdoor monocular visionaided navigation and control of unmanned aircraft Jour-nal of Field Robotics 30(3) 415ndash438

Dryanovski I Valenti R G amp Xiao J (2013) An open-sourcenavigation system for micro aerial vehicles AutonomousRobots 34(3) 177ndash188

Eich M amp Vogele T (2011) Design and control of a lightweightmagnetic climbing robot for vessel inspection In IEEE 19thMediterranean Conference on Control and Automation(pp 1200ndash1205) Corfu Greece

Fletcher J amp Kattan R (2009) A ballast tank coating in-spection data management system httpwwwelcoshipcompapershtm

Fraundorfer F Heng L Honegger D Lee G H Meier LTanskanen P amp Pollefeys M (2012) Vision-based au-tonomous mapping and exploration using a quadrotorMAV In IEEERSJ International Conference on IntelligentRobots and Systems (IROS) (pp 4557ndash4564) VilamouraAlgarve Portugal

Grzonka S Grisetti G amp Burgard W (2012) A fully au-tonomous indoor quadrotor Transactions on Robotics28(1) 90ndash100

Gurdan D Stumpf J Achtelik M Doth K-M HirzingerG amp Rus D (2007) Energy-efficient autonomous four-rotor flying robot controlled at 1 kHz In IEEE InternationalConference on Robotics and Automation (ICRA) (pp 361ndash366) Roma Italy

Haralick R amp Shapiro L (1992) Computer and robot visionReading MA Addison-Wesley

Isard M amp Blake A (1998) CONDENSATIONmdashConditionaldensity propagation for visual tracking International Jour-nal of Computer Vision 29 5ndash28

Jetstream (2013) Jetstream Magnetic Crawler M250 httpwwwjetstreameuropecoukproductsmagnetic-crawlerhtml

Jung M Schmidt D amp Berns K (2010) Behavior-based ob-stacle detection and avoidance system for the omnidi-rectional wall-climbing robot CROMSCI In Fujimoto HTokhi M O Mochiyama G S Virk G S (eds) Emerg-ing trends in mobile robotics (pp 73ndash80) Singapore WorldScientific

Kaess M Johannsson H Englot B Hover F S amp LeonardJ J (2010 May) Towards autonomous ship hull in-spection using the Bluefin HAUV In Ninth Interna-tional Symposium on Technology and the Mine ProblemMonterey CA Available at httpwwwccgatechedusimkaesspubKaess10istmphtml

Kalra L Guf J amp Meng M (2006) A wall climbing robot foroil tank inspection In IEEE International Conference onRobotics and Biomimetics (IEEE-ROBIO) (pp 1523ndash1528)Kunming China

Kim H Seo K Lee K amp Kim J (2010) Development of amulti-body wall climbing robot with tracked wheel mech-anism In Fujimoto H Tokhi M O Mochiyama G SVirk G S (eds) Emerging trends in mobile robotics (pp439ndash446) Singapore World Scientific



Koveos Y Kolyvas T amp Drikos L (2012) Robotic armdevelopment for ultrasonic inspection In InternationalConference on Computers and IT Applications in theMaritime Industries (COMPIT) (pp 258ndash268) LiegeBelgium

Matsue A Hirosue W Tokutake H Sunada S amp OhkuraA (2005) Navigation of small and lightweight helicopterTransactions of the Japan Society for Aeronautical andSpace Sciences 48(161) 177ndash179

Meier L Tanskanen P Heng L Lee G H FraundorferF amp Pollefeys M (2012) PIXHAWK A micro aerialvehicle design for autonomous flight using onboard com-puter vision Autonomous Robots 33(1-2) 21ndash39

Miko (2013) Steel-Climber httpwwwmikonosteel-climber

Ortiz A Bonnin F Gibbins A Apostolopoulou P Bate-man W Eich M Spadoni F Caccia M amp Drikos L(2010) First steps towards a roboticized visual inspectionsystem for vessels In International Conference on Emerg-ing Technologies and Factory Automation (ETFA 2010)Bilbao Spain

Ortiz F Pastor J A Alvarez B Iborra A Ortega N Ro-driguez D amp Conesa C (2007) Robots for hull shipcleaning In 15th IEEE International Symposium onIndustrial Electronics (ISIE) (pp 2077ndash2082) VigoSpain

Raut S Patwardhan A amp Henry R (2010) Design and con-struction of glossy surface climbing biped robot In Fu-jimoto H Tokhi M O Mochiyama H amp Virk G S(eds) Emerging trends in mobile robotics (pp 423ndash430)Singapore World Scientific

Roberts J F Stirling T Zufferey J-C amp Floreano D (2007)Quadrotor using minimal sensing for autonomous indoor

flight In European Micro Air Vehicle Conference andFlight Competition (EMAV) Toulouse France

Shang J Bridge B Sattar T Mondal S amp Brenner A (2008)Development of a climbing robot for inspection of longweld lines Industrial Robot An International Journal35(3) 217ndash223

Silva M F amp Tenreiro J (2010) A survey of technologies andapplications for climbing robots locomotion and adhesionIn Behnam M (ed) Climbing and walking robots (Chap1 pp 1ndash22) Rijeka InTech

Tache F Fischer W Caprari G Siegwart R Moser R ampMondada F (2009) Magnebike A magnetic wheeled robotwith high mobility for inspecting complex-shaped struc-tures Journal of Field Robotics 26(5) 453ndash476

Tanneberger K amp Grasso A (2011) MINOAS deliverable(D1) Definition of the inspection plandefinition ofacceptance criteria httpminoasprojecteuexcibitionPublicationsPDFD1pdf

Thrun S Fox D Burgard W amp Dellaert F (2000) RobustMonte Carlo localization for mobile robots Artificial In-telligence 128(1-2) 99ndash141

Wenzel K E Masselli A amp Zell A (2011) Automatic takeoff tracking and landing of a miniature UAV on a movingcarrier vehicle Journal of Intelligent Robotic Systems 61(1-4) 221ndash238

Yamaguchi T amp Hashimoto S (2010) Fast crack detec-tion method for large-size concrete surface images usingpercolation-based image processing Machine Vision andApplications 21(5) 797ndash809

Zingg S Scaramuzza D Weiss S amp Siegwart R (2010May) MAV navigation through indoor corridors usingoptical flow In International Conference on Robotics andAutomation (pp 3361ndash3368) Anchorage AK


本文献由ldquo学霸图书馆-文献云下载rdquo收集自网络仅供学习交流使用

学霸图书馆（wwwxuebalibcom）是一个ldquo整合众多图书馆数据库资源

提供一站式文献检索和下载服务rdquo的24 小时在线不限IP

图书馆

图书馆致力于便利促进学习与科研提供最强文献下载服务

图书馆导航

图书馆首页文献云下载图书馆入口外文数据库大全疑难文献辅助工具


Figure 1 Illustration of traditional inspection methods (left) general cargo ship (center) staging previous to inspection and(right) cherry-picking [Source (leftcenter) Lloydrsquos Register (right) httpwwwstandard-clubcom]

a bulk carrier is the most time-consuming part of the inspec-tion process As can be seen in Figure 1 (center) traditionalship surveying methods include prior to the survey theinstallation of scaffolding to allow the surveyor to inspectstructures such as bulkheads beams stiffeners and brack-ets which are usually several meters above the bottom of thehold Besides the scaffolding ldquocherry-pickingrdquo methods arealso employed in this case the surveyor reaches the point ofinterest inside a basket transported by a tower crane or bya hydraulic arm [Figure 1 (right)] Clearly these procedurescan be dangerous for the surveyor For this reason and be-cause of the high costs of gaining access to a ship for inspec-tion the EU-funded research project MINOAS (Caccia et al2010b) set up a consortium to introduce robots into the shipsurveying process The basis of the consortiumrsquos expertisecomprised two classification societies different marine ser-vice providers and partners involved in robotics researchThe key idea was to introduce a set of novel tools to enhancethe ship surveying process For the interested reader a moredetailed discussion of the application scenario can be foundin Ortiz et al (2010)

This paper reports the results of the introduction ofheterogeneous robots to the area of close-up surveys of thestructural elements of large-tonnage vessels where most ofthe work is still performed manually The effort is a mix-ture of novelty and integration and the proportion of eachis different for every platform Nevertheless the main con-tribution is the fully integrated inspection system cover-ing all the stages of an inspection procedure based on theuse of robots and supporting software something that didnot exist before the project MINOAS We introduce a het-erogeneous robot team with different locomotion abilitiesnamely a micro-aerial vehicle (MAV) a lightweight mag-netic crawler supported by an external positioning unitand a heavyweight magnetic crawler equipped with a ma-nipulator arm for thickness measurement Two additionalsystems were developed to assist the surveyor a spatial con-tent management system (SCMS) to host and present in acentralized way the inspection data collected by all the plat-forms and a visual defect detection solution for automatic

defect detection They are all described in the followingsections

2 REENGINEERED INSPECTION PROCEDURE

On the basis of the expertise of the classes involved andthe maritime industry and as part of the MINOAS projectworking plan a total of three stages were defined in or-der to implement an inspection procedure based on robotsand compatible with the requirements of the classes (Tan-neberger amp Grasso 2011)

Stage 1 Fast visual inspection overview The goal of thisstage is to cover large areas of selected parts of the inner hullsupplying visual inspection data to get a rough overviewabout the state of the vessel and searching for defects suchas coating breakdown corrosion and cracks The imagescollected must be tagged with pose information since theareas of interest for Stages 2 and 3 are visually selectedin this phase Due to the aforementioned requirement anaerial platform turns out to be the best option In more detailthe vehicle must be able to perform vertical stationary andlow-speed flight in indoor environments

Stage 2 Visual close-up survey The procedure in Stage 2is to get a better impression of the coating and the possibledamages and mark the defective areas in order to repairthe damage or perform thickness measurements The vehi-cle must be capable of moving on horizontal sloped andvertical ferromagnetic surfaces At this stage the camera ofthe robot has to be steady and very close to the structuralparts in order to provide high-resolution high-quality im-ages Proper lighting of the area under inspection shouldbe provided Marking the defects directly on the struc-ture once they are confirmed by the surveyor would be anoptional feature of the robot The acquired images haveto be tagged with positioning information Finally ease ofsystem setup is also required Under these constraints alightweight magnetic crawler was considered to be a suit-able platform

Stage 3 Thickness measurement collection At selectedparts identified during Stage 2 the thickness of the material









times






















































Bel(xt ) = μ1

σradic

2πeminus 1


timesint


(1)




σradic





























ylowastn = 1

a0

⎛⎝ Psum

i=0


j=1

ajy1nminusj

⎞⎠ n = 0 N


yn = 1a0

⎛⎝ Psum

i=0


Qsumj=1

ajynminusj

⎞⎠ n = 0 N




thickness = c

2dt

(arg max

j

sumn

xnxnminusj

) j isin nmin nmax










i=0

sum31j=0 p(i j )2



42 Crack Detection










ship

cargo

port

hold1

star-board port

hold2

star-board

bridge

super-structure

crane1




























MaxVertical Max



clean 451 198 block










11

12

13

14

15

16

17

18

19

3 4 5 6 7 8 9 10 11 12 13 14

Pos

ition

Y (

m)

Position X (m)

measurement error







56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)


0

1

2

3

4

Y (m)

Z (

m)


0 1000 2000 3000 4000 5000 600005

1


r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]





















(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航









times






















































Bel(xt ) = μ1

σradic

2πeminus 1


timesint


(1)




σradic





























ylowastn = 1

a0

⎛⎝ Psum

i=0


j=1

ajy1nminusj

⎞⎠ n = 0 N


yn = 1a0

⎛⎝ Psum

i=0


Qsumj=1

ajynminusj

⎞⎠ n = 0 N




thickness = c

2dt

(arg max

j

sumn

xnxnminusj

) j isin nmin nmax










i=0

sum31j=0 p(i j )2



42 Crack Detection










ship

cargo

port

hold1

star-board port

hold2

star-board

bridge

super-structure

crane1




























MaxVertical Max



clean 451 198 block










11

12

13

14

15

16

17

18

19

3 4 5 6 7 8 9 10 11 12 13 14

Pos

ition

Y (

m)

Position X (m)

measurement error







56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)


0

1

2

3

4

Y (m)

Z (

m)


0 1000 2000 3000 4000 5000 600005

1


r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]





















(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航













































Bel(xt ) = μ1

σradic

2πeminus 1


timesint


(1)




σradic





























ylowastn = 1

a0

⎛⎝ Psum

i=0


j=1

ajy1nminusj

⎞⎠ n = 0 N


yn = 1a0

⎛⎝ Psum

i=0


Qsumj=1

ajynminusj

⎞⎠ n = 0 N




thickness = c

2dt

(arg max

j

sumn

xnxnminusj

) j isin nmin nmax










i=0

sum31j=0 p(i j )2



42 Crack Detection










ship

cargo

port

hold1

star-board port

hold2

star-board

bridge

super-structure

crane1




























MaxVertical Max



clean 451 198 block










11

12

13

14

15

16

17

18

19

3 4 5 6 7 8 9 10 11 12 13 14

Pos

ition

Y (

m)

Position X (m)

measurement error







56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)


0

1

2

3

4

Y (m)

Z (

m)


0 1000 2000 3000 4000 5000 600005

1


r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]





















(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航






















ylowastn = 1

a0

⎛⎝ Psum

i=0


j=1

ajy1nminusj

⎞⎠ n = 0 N


yn = 1a0

⎛⎝ Psum

i=0


Qsumj=1

ajynminusj

⎞⎠ n = 0 N




thickness = c

2dt

(arg max

j

sumn

xnxnminusj

) j isin nmin nmax










i=0

sum31j=0 p(i j )2



42 Crack Detection










ship

cargo

port

hold1

star-board port

hold2

star-board

bridge

super-structure

crane1




























MaxVertical Max



clean 451 198 block










11

12

13

14

15

16

17

18

19

3 4 5 6 7 8 9 10 11 12 13 14

Pos

ition

Y (

m)

Position X (m)

measurement error







56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)


0

1

2

3

4

Y (m)

Z (

m)


0 1000 2000 3000 4000 5000 600005

1


r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]





















(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航




i=0

sum31j=0 p(i j )2



42 Crack Detection










ship

cargo

port

hold1

star-board port

hold2

star-board

bridge

super-structure

crane1




























MaxVertical Max



clean 451 198 block










11

12

13

14

15

16

17

18

19

3 4 5 6 7 8 9 10 11 12 13 14

Pos

ition

Y (

m)

Position X (m)

measurement error







56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)


0

1

2

3

4

Y (m)

Z (

m)


0 1000 2000 3000 4000 5000 600005

1


r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]





















(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航



ship

cargo

port

hold1

star-board port

hold2

star-board

bridge

super-structure

crane1




























MaxVertical Max



clean 451 198 block










11

12

13

14

15

16

17

18

19

3 4 5 6 7 8 9 10 11 12 13 14

Pos

ition

Y (

m)

Position X (m)

measurement error







56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)


0

1

2

3

4

Y (m)

Z (

m)


0 1000 2000 3000 4000 5000 600005

1


r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]





















(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航






















MaxVertical Max



clean 451 198 block










11

12

13

14

15

16

17

18

19

3 4 5 6 7 8 9 10 11 12 13 14

Pos

ition

Y (

m)

Position X (m)

measurement error







56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)


0

1

2

3

4

Y (m)

Z (

m)


0 1000 2000 3000 4000 5000 600005

1


r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]





















(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航




11

12

13

14

15

16

17

18

19

3 4 5 6 7 8 9 10 11 12 13 14

Pos

ition

Y (

m)

Position X (m)

measurement error







56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)


0

1

2

3

4

Y (m)

Z (

m)


0 1000 2000 3000 4000 5000 600005

1


r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]





















(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航



56

7 01

23

minus2

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)

0 1 2 3minus2

minus1

0

1

2

3

4

Y (m)

Z (

m)

45

67 minus5

minus4minus3

minus2minus1

0

minus1

0

1

2

3

4

Y (m)X (m)

Z (

m)


0

1

2

3

4

Y (m)

Z (

m)


0 1000 2000 3000 4000 5000 600005

1


r 1r3d

[m] minus

(g

rk)

0 1000 2000 3000 4000 5000 6000minus05

0

05

e d [m]

0 1000 2000 3000 4000 5000 6000minus20

0

20

eψ

e ψ [d

eg]

[s10]





















(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航














(a)

















ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航








ACKNOWLEDGMENTS





REFERENCES
















































图书馆


图书馆导航




REFERENCES
















































图书馆


图书馆导航
























图书馆


图书馆导航


a robot application for marine vessel inspectiondownload.xuebalib.com/xuebalib.com.36861.pdf ·...

Documents