toward automated power line corridor monitoring

Toward Automated Power Line Corridor MonitoringUsing Advanced Aircraft Control and MultisourceFeature Fusion

• • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • •

Zhengrong Li, Troy S. Bruggemann, Jason J. Ford, Luis Mejias, and Yuee LiuAustralian Research Centre for Aerospace Automation, Queensland University of Technology, Boronia Road, Eagle Farm,Queensland 4009, Australiae-mail: [email protected], [email protected]

Received 14 February 2011; accepted 6 September 2011

The conventional manual power line corridor inspection processes that are used by most energy utilities arelabor intensive, time consuming, and expensive. Remote sensing technologies represent an attractive and cost-effective alternative approach to these monitoring activities. This paper presents a comprehensive investigationinto automated remote sensing–based power line corridor monitoring, focusing on recent innovations in thearea of increased automation of fixed-wing platforms for aerial data collection and automated data processingfor object recognition using a feature fusion process. Airborne automation is achieved by using a novel ap-proach that provides improved lateral control for tracking corridors and automatic real-time dynamic turningfor flying between corridor segments; we call this approach PTAGS (Power line Tracking Automatic GuidanceSystem). Improved object recognition is achieved by fusing information from multisensor (LiDAR and im-agery) data and multiple visual feature descriptors (color and texture). The results from our experiments andfield survey illustrate the effectiveness of the proposed aircraft control and feature fusion approaches. C© 2011Wiley Periodicals, Inc.

1. INTRODUCTIONThere is no doubt about the increased reliance that mod-ern societies have on electricity. Electrical companies areunder continuous and significant pressure to ensure reli-able supply and distribution of electricity. Whereas trees,shrubs, and other vegetation are of remarkable importanceto the environment and our daily life, inappropriate vege-tation around power lines can represent a significant riskto public safety and is one of the main causes of poweroutages. It is commonly known that trees falling acrosspower lines are the largest cause of power failures, causingwidespread power outages and bushfires (Ituen & Sohn,2010; Mills et al., 2010). Hence, it is not surprising thatpower line corridor vegetation management procedureshave become a significant maintenance cost for electricalcompanies. For example, Ergon Energy, Australia’s largestgeographic footprint energy distributor, currently spendsmore than $80 million a year inspecting and managing veg-etation encroachments on power lines. Correct and efficientvegetation management not only reduces the overall costbut also aids in continuous electricity supply by preventingdamage to power lines through removal of intrusive trees.Ineffective procedures can result in the loss of reliability inelectricity transmission, produce serious hazards, and ex-pose electrical companies to significant financial penalties.

A power line corridor describes the strip of land uponwhich utility companies construct their electrical infra-

structure. Monitoring power line corridors is crucial forthe reliability of electricity transmission. Trees and shrubsoften create obstructions in corridors and pose risks topower lines, and therefore utility companies need to scru-tinize where and how trees grow in or close to powerline corridors (Ituen & Sohn 2010). In urban areas, veg-etation encroachment is less serious than in rural areasas access is much easier and prompt maintenance can beachieved. Moreover, local councils and private land own-ers regularly maintain their trees, facilitating the overallmaintenance process. However, in rural areas, inspectionmaintenance becomes difficult due to limited access andlarge distances to cover. In these areas, traditional calendar-based tree trimming is often a strategy used by energy dis-tributors. Other short-term strategies might be to identifyand remove nearby objects (i.e., buildings and vegetation)found near power lines. More generally, the risk of man-made structures can be controlled through building regu-lations. However, vegetation grows naturally, and particu-larly in rural areas, the growth of vegetation is unmanaged.Strong winds and storms can bring branches or even en-tire trees into contact with power lines. Unmanaged vegeta-tion can also grow up into power lines and cause bushfires.Unfortunately, vegetation management over large powerline networks is cumbersome and expensive. To managethe trade-off between safety and cost, utility companiesoften introduce a regular maintenance cycle (say, once

Journal of Field Robotics 29(1), 4–24 (2012) C© 2011 Wiley Periodicals, Inc.View this article online at wileyonlinelibrary.com • DOI: 10.1002/rob.20424

Li et al.: Aerial Robotics Application • 5

every 5 years). This strategy assumes that once inspected,properly maintained vegetation can be assumed to remainseparated from the power line infrastructure until the main-tenance cycle is repeated 5 years later. However, trees of-ten grow unexpectedly during the period between main-tenance due to inaccurate vegetation growth modelingand climate changes. For example, the Australian state ofQueensland is subject to extreme weather conditions, rang-ing from drought to cyclones. Moreover, Queensland cansuffer extended periods of dry conditions, which signifi-cantly increase the risk of fire. Strong winds and water-logged ground can result in trees falling across and bring-ing down power lines, especially when inappropriate veg-etation species have grown too close to power lines. Con-ventional vegetation management strategies consider onlythe direct clearance of a small number of trees within tightcorridors below the power line infrastructure, whereas thegrowth of nearby and potentially troublesome vegetationthat is slightly outside the corridor is often neglected. Al-though indiscriminate removal of all vegetation within apower line corridor is neither cost effective nor ecofriendly,better longer term strategies can be achieved by improvedmonitoring and modeling of the growth of all vegetationsurrounding power line infrastructure.

The subjective nature of conventional maintenancestrategies often results in some zones being trimmed morefrequently than required or conversely, some zones not be-ing trimmed often enough. Remote sensing technologiesrepresent an attractive and potentially automated solutionfor power line corridor monitoring activities. Recent effortstoward remote sensing–based methods include improveddata collection using satellite sensors (Beltrame, Jardini,Jacbsen, & Quintanilha, 2007; Kobayashi, Karady, Heydt,& Olsen, 2009), an airborne stereo vision system (Sunet al., 2006), and unmanned aerial vehicles (UAVs) (Jones,Golightly, Roberts, Usher, & Earp, 2005; Li, Liu, Walker,Hayward, & Zhang, 2010). Recently airborne laser scan-ning has attracted particular attention in power line corri-dor monitoring due to the possibility of achieving three-dimensional (3D) models of infrastructure (Chaput, 2008;Jwa, Sohn, and Kim, 2009; Lu & Kieloch, 2008). Many, ifnot most, utility companies and researchers use commer-cial data and software but also generate algorithms tai-lored to their needs. However, there are some critical issuesto be addressed regarding automated data collection andprocessing over complicated power line networks. For ex-ample, the quality of collected data has much to do withaircraft platform stability. The complex nature of navigat-ing aircraft over extensive numbers of power line networkscalls for an increased level of flight automation. To increasethe reliability of the information extraction from remotesensing data, combining the complementary informationderived from multisource data can be very useful. The ef-fective fusion of multisource data provides the opportu-nity for more robust operational performance and decisionmaking in power line corridor monitoring.

This paper begins in Section 2 with a survey of re-mote sensing–based power line corridor monitoring. Wethen detail two important aspects of the power line cor-ridor monitoring problem: advanced aircraft control forpower line corridor monitoring and multisensor data fu-sion for vegetation classification. For this purpose, wefirst summarize our receding virtual waypoint and pre-cision guidance tracking approach reported previously inBruggemann, Ford, and Walker (2010). We then summarizeour automated data processing work reported previouslyin Li, Hayward, Walker, and Liu (2011); Li, Liu, et al. (2010),and Mills, et al. (2010). In Section 3, we present some newflight tests results that evaluate the ability of the inspec-tion aircraft to maintain a LiDAR (light detection and rang-ing) swath width over the features of interest. Then, a newautomated aircraft behavior for maneuvering capability atpower line corridor corners is described and flight test re-sults presented. In Section 4, new approaches and resultsfor the combined use of LiDAR and multispectral imagedata, as well as the fusion of multiple visual feature descrip-tors, are presented. The paper ends in Section 5 by present-ing some conclusions drawn from the lessons we learnedduring our 3-year power line corridor monitoring project.

2. TECHNOLOGY OVERVIEW

2.1. Power Line Corridor Monitoring Using AerialRemote Sensing

Ergon Energy has a long-term strategy of managing vege-tation according to different species; species can be gener-ally categorized as either desirable or undesirable. Specieswith fast growth rates and that also have the potential toreach a mature height of more than 4 m are defined as unde-sirable. These undesirable species often pose high risks toelectrical infrastructure and therefore should be identifiedand removed. It is also worth mentioning that a reason-able long-term maintenance strategy is to encourage low-growing trees or shrubs because they are expected to com-pete with tall-growing species and deprive the taller trees oflight and nutrients. These low-growing species, along withthe rare and endangered species, are defined as desirablespecies that should be managed differently.

The need to manage risk motivates the collection ofdata over power line corridors and the identification ofobjects of interest in order to assess risk levels and guidefield workers for vegetation clearance in the corridors. Re-mote sensing represents a particularly attractive solutionfor power line corridor monitoring. Actually, aerial ve-hicles have been intensively used in power line inspec-tion for a long time. One present practice is to fly he-licopters/airplanes along the corridor and try to iden-tify dangerous trees and assess the condition of overheadlines assets by visual observation. Such visual inspection istime consuming and labor intensive. The advances in sen-sor technologies and intelligent computing techniques pro-vide the opportunity to move from traditional vegetation

Journal of Field Robotics DOI 10.1002/rob

6 • Journal of Field Robotics—2012

management strategies to more automated, accurate, andcost-effective solutions.

2.1.1. Remote Sensing Platforms

Satellites and aircraft are the most widely used platformsfor remote sensing in Earth-Observing data collection. Cur-rent satellite sensors are not the best choice for monitoringpower line corridors due to two critical limitations: the un-favorable revisit time and lack of choices in optimum spa-tial and spectral resolutions. At the most practical level,most collections of data gathered from satellites are avail-able only on predetermined schedules, and even those withan “on-demand” capability are also limited by their orbitsand the demands of other users. In contrast, airborne datacollection offers a much greater level of flexibility. Anotheradvantage of an airborne platform is that different sensorpayloads can be easily fitted. Whereas the sensors launchedon a satellite are rarely changeable. As a consequence, air-borne systems can be regularly upgraded as sensor technol-ogy advances. Improvements to sensors include systemswith higher spectral and spatial resolution and advancedmicrowave or LiDAR sensors. In addition, higher spatialresolutions are easier to obtain from airborne platforms,due to their low altitude. A limitation that impedes large-scale airborne remote sensing applications is that the tradi-tional piloted airborne platforms involve high operationalcosts. Moreover, using piloted aircraft for power line in-spection will place the pilots at a greater level of risk. Many,if not most, airborne LiDAR systems use helicopter plat-forms (Ituen & Sohn 2010). Although flying LiDAR witha rotorcraft has some advantages over fixed-wing aircraftwhere tight turns are required, a fixed-wing aircraft has ad-vantages over rotorcraft in terms of cost per kilometer sur-veyed in the large-scale aerial inspection tasks that we areconsidering in this paper.

Remote sensors mounted on UAVs could fill this ca-pability gap, providing a cheap and flexible way to gatherspatial data from power line corridors that can also meetthe requirements of spatial, spectral, and temporal reso-lutions. Recent developments in the aerial vehicles them-selves and associated sensing systems make UAV plat-forms increasingly attractive for both research and oper-ational mapping (Berni, Zarco-Tejada, Suarez, & Fereres,2009; Gurtner, et al., 2009). One of the main barriers tousing UAVs is their inability to carry power-demandingand heavy payloads. LiDAR systems are usually too heavyfor small/medium-sized UAV platforms. This UAV limi-tation may be overcome in the near future as there arealready small LiDAR systems in the market suitable forUAVs. However, the performance of these units in termsof the quality of data collected is currently well away fromtheir full-sized counterparts, and more development is re-quired. Another key barrier to using UAVs is the aviationregulatory issues associated with using UAVs outside of the

visual range of an operator. However, if both these barri-ers can be addressed, the combination of small LiDAR-typesystems and advanced UAVs might represent a powerfulaerial inspection technology.

The above-mentioned limitations of current UAVtechnology motivate consideration of manned semiau-tonomous systems as a technological step from manuallypiloted inspection toward fully autonomous operations.Manned semiautonomous systems would provide in-creased automation in the aerial survey task but also al-low the pilot to remain onboard to monitor and providehuman control and oversight of the flight. Manned semi-autonomous operation has advantages over the currentmanually piloted approach such as reducing the crewingrequirements (pilot only, instead of sensor operator and pi-lot) and allowing a pilot without specialized skills in fly-ing above power lines to safely and routinely conduct suc-cessful aerial surveys. Safety could also be improved byreducing pilot fatigue and by allowing the pilot to focuson important aircraft safety-of-operation tasks such as the“see-and-avoid” function.

2.1.2. Automated Data Processing

There are two kinds of remote sensing: passive remotesensing and active remote sensing. Passive sensors detectnatural radiation that is emitted or reflected by the objector surrounding area being observed. Reflected sunlight isthe most common source of radiation measured by pas-sive sensors. Optical remote sensing images such as satel-lite and airborne multispectral imagery are collected frompassive sensors. The spatial resolution and spectral reso-lution are the two most important characteristics of opti-cal remote sensing imagery. Spatial resolution, commonlyreferred to as “pixel size” in digital images, has a closerelationship with the information content that can be ex-tracted from the image; however, higher spatial resolutionis not always beneficial. In many sensor processing prob-lems, images with a spatial resolution near the size of theobject of interest are often preferred (Lefsky & Cohen 2003).Spectral resolution is the richness of spectral informationin optical remote sensing imagery. Different materials re-flect and absorb differently at different wavelengths. Spec-tral features are the specific combination of reflected andabsorbed electromagnetic radiation at varying wavelengthswhich can uniquely identify an object. For example, near-infrared (NIR) wavelengths have been successful applied toestimating vegetation biophysical properties (Rautiainen,2005). Active sensors, on the other hand, emit energy inorder to scan objects and areas, whereupon a passive sen-sor then detects and measures the radiation that is reflectedor backscattered from the target. LiDAR is an example ofactive remote sensing in which the properties of returnedscattered light are used to find range or other informationof a distant target. Traditionally, the primary use of LiDAR



data is to obtain altitude data and generate digital terrainmodels (DTM). In recent years, however, the range of ap-plications in which laser scanning can be used has greatlybroadened. With the advancement of sensor technology, theachievable resolution of point clouds makes it possible tomap individual trees and power lines from airborne laserscanning data.

As may already be evident from the preceding sec-tion, aerial remote sensing represents an attractive solutionfor power line corridor mapping. Compared to corridormapping, automated and intelligent information extrac-tion from remotely sensed data is even more challeng-ing. One special need for power line corridor monitoringis to automatically detect the objects of interest for fur-ther interpretation and decision making (major objects ofinterest include power line assets and vegetation). Auto-mated data processing aims to automatically detect theseobjects from aerial imagery and tries to extract more spe-cific information such as vegetation species and heightinformation.

Risk assessment of power lines and adjacent trees ismeaningful only when power lines and trees can be de-tected. A number of papers have been published on powerline detection during the past few years, from both two-dimensional (2D) imagery and 3D LiDAR point cloud data.A straight line can approximate a power line segment inaerial images. Therefore, some classic line detection algo-rithms such as the Hough transform may be used to de-tect power lines in images. Yan et al. employed a Radontransform to extract line segments of the power lines, fol-lowed by a grouping method to link each segment and aKalman filter to connect the segments into an entire line(Yan, Li, Zhou, Zhang, & Li, 2007). Li et al. developed afilter based on a simplified pulse coupled neural networkmodel (Li, Liu, et al., 2010). This filter can simultaneouslyremove the background noise and generate edge maps. Af-ter that, an improved Hough transform is used by per-forming knowledge-based line clustering in Hough spaceto refine the detection results. The accuracy of a 2D image–based power line detection algorithm depends on the qual-ity of the image. Low spatial resolution and motion blur arethe two major causes of the failure of the power line detec-tion algorithms. LiDAR is more popular for power line sur-veys than image-based approaches because it can providehigh-density point cloud data and does not rely on illumi-nation conditions. LiDAR can more effectively generate ac-curate elevation and terrain models, which can also help toremove terrain points and other similar linear features (e.g.,fences). Moreover, the 3D nature of LiDAR data makes itpossible to model the sag of lines, which is also crucial forthe maintenance of electrical infrastructure. For 3D LiDARpoint cloud data, line detection can be conducted either byclustering similar features in a voxel (Jwa et al., 2009) ormapping 3D data to two 2D planes (horizontal and verti-

cal) and then roughly detecting the power line points in thehorizontal plane and reconstructing the catenary curve inthe vertical plane (Liu, Li, Hayward, Walker, & Jin, 2009).No matter which method is used, prior removal of non–power line points (e.g., terrain, trees, and buildings) willalways be helpful to reduce the false-positive rate in powerline detection.

Tree detection from remote sensing imagery is wellresearched, particularly in the context of forest and plan-tation management (Mallinis, Koutsias, Tsakiri-Strati, &Karteris, 2008; Pouliot, King, Bell, & Pitt, 2002). Accord-ing to our literature review, image-based tree detectionmethods can be broadly categorized as either local max-ima/minima, template matching, region growing, or edgedetection approaches (Li, Hayward, Zhang, Liu, 2008). Asthe most widely used tree detection method, the local max-ima/minima approach uses the following assumption: theradiometric properties of a tree can be described througha mountainous landscape in which peaks are approximatecrown apexes and surrounding valleys represent the spacebetween crowns or where crowns overlap or touch (Pouliot,King, & Pitt 2005). However, this assumption is not al-ways true. In some real situations, multiple trees touchclosely and have no distinct dark boundary between treecrowns. Research indicates that in this case template match-ing gives better results; however it is very time consuming(Erikson & Olofsson, 2005). Another drawback of the localmaxima/minima approach is the initial assumptions aboutcrown size and shape and the relative inflexibility of themodel to accommodate irregular crown form. Consideringthe spectral properties of vegetation, particular NIR bandcan be very helpful to detect trees. Li et al. employed asimplified pulse-coupled neural network (PCNN) that usesspectral features as input, postprocessed using morpholog-ical reconstruction (Li, Hayward, Zhang, Liu, & Walker,2009). The algorithm has been shown to outperform bothJSEG (Deng & Manjunath, 2001) and TreeAnalysis (Erikson,2003) in tree crown segmentation, but the primary errorsource is the undersegmentation of tree clusters due to thecrown overlap. LiDAR system can measure both verticaland horizontal structures of object, which makes it a goodmeans for detecting individual trees and estimating treeparameters. Many LiDAR-based tree detection algorithmsborrow the idea from image-based methods. Leckie et al.(2003) applied the valley-following method, and Kwak,Lee, Lee, Biging, and Gong (2007) segmented individualtrees from LiDAR data using extended maxima transfor-mation of the morphological image-analysis method. De-spite the benefit of 3D information in tree segmentation, thelack of spectral and texture information makes it hard to de-rive more detailed vegetation information (e.g., biophysicalproperties and species) only from LiDAR data. Combina-tion of LiDAR data and optical imagery is considered to bea very promising solution.



Display (and / or operator) guidance to pilot

Pilot manual control

Aircraft Performance

Pilot’s senses (visual & cockpit display)

Figure 1. Manual control of inspection aircraft (current industry practice).

2.2. Control of Airborne Platform

Maintaining stability and control of a fixed-wing airborneplatform above a power line corridor is necessary for re-liable data capture of ground assets and features such aspower lines and vegetation. To achieve data capture ob-jectives with a body-fixed, downward-pointing LiDAR orcamera, the aircraft must maintain an accurate track overthe power line corridor, with minimal lateral position andangular heading deviation from the desired flight path overthe line. The lateral position deviation is referred to ascross-track error and is a key parameter that the controllerattempts to minimize. At the same time, platform orien-tation must be controlled and stabilized because excessiveaircraft bank angle may result in parts of the corridor beingmissed or poor captured data quality (Bruggemann et al.,2010; Holt & Beard, 2010; Nelson, Barber, McLain, & Beard,2006).

2.2.1. Limitations of Current Aerial Survey IndustryPractices and Technologies for Inspecting Power LineCorridors

Currently in the aerial survey industry, platform stabilityand tracking is typically maintained by manual pilot con-trol with operator assistance. Pilots are required to concen-trate on a flight display for extended periods of time dur-ing inspection and must be allowed regular breaks for restand recovery. A representative diagram of this approach isgiven in Figure 1. However, the power line inspection taskcan involve flying extensive amounts of power line corri-dor on a routine basis. For example, Ergon Energy’s net-work in Queensland, Australia, consists of 150,000 km ofpower line and could require 4,000 h or more of flying timeto inspect the whole network. Flying aircraft at low altitudeover power lines for up to 4 or 5 h, at a time is a tediousand potentially dangerous task under manual piloted con-trol. We argue that improved safety of airborne operationsand an increased reliability of data capture may be achievedthrough the increased use of flight automation for inspec-tion tasks.

However, our study of current autopilots and GPS nav-igators showed that these existing technological solutionsare not designed for controlling an aircraft to track powerlines—they are designed chiefly for routine navigation in-cluding waypoint-to-waypoint navigation and conducting

holding patterns and procedure turns. Further, standardautopilots and GPS navigators typically lack the configura-bility and tuneability required to allow them to performin a suitable way for power line corridor inspection. Un-like waypoint-to-waypoint navigation, tracking power linecorridors for inspection requires both aircraft translationaland rotational motion to be controlled (Bruggemann et al.,2010). It should be clear from consideration of the powerline network to be flown (Figure 2) that, from an automaticcontrol point of view, a solution is required different fromthat provided by standard autopilots and navigators, be-cause of the following:

• The power line network consists of many large and sud-den changes in line direction as well as short spacingbetween lines sections. This presents automatic controlchallenges such as how to maintain accurate positionover the line, yet also limit and control aircraft orienta-tion (e.g., bank angle) such that no sections of the cor-ridor are missed (Bruggemann et al., 2010). Also, ter-rain variations and aircraft altitude and speed need to beconsidered.

Figure 2. A 20 × 15 km area of power line network (blacklines). The distribution, size, and orientation of lines indicatechallenges for automated guidance and control of a fixed-wingairborne platform. Large and sudden changes in line direc-tion and short spaces between lines sections require specializedcontrol techniques.



• A very large number of waypoints are required to spec-ify a power line network under inspection. This presentschallenges in both waypoint database management andflight planning. For example, Ergon’s electrical distribu-tion in Queensland is described by approximately 1 mil-lion waypoints.

• Typically, the total width of a specified power line corri-dor must be inspected (not just the power line itself), andthis requirement necessitates consideration and moni-toring of aircraft orientation and the sensor footprint onthe ground.

• Automatic turn replanning around power line corridorsegments (in cases of excessive cross-track error or bankangle requiring a go-around maneuver) is also advanta-geous for reliability of data capture.

New automatic control technology could allow im-proved safety by removing the need for the pilot to man-ually maintain track and orientation over the line andallowing the pilot to focus on operating the aircraft and sit-uational awareness. It could also facilitate single-crew op-erations that do not require an additional operator to assistthe pilot.

Past research has focused on the problem of automaticcontrol of airborne platforms for the purpose of aerial in-spection of linear assets such as rivers, roads, pipelines,and power lines (Egbert & Beard, 2007; Frew, et al.,2004; Holt & Beard, 2010; Rathinam, Kim, Soghikian, &Sengupta, 2005). Early studies discovered that simpleproportional–integral–derivative (PID)-based control loopscan lead to poor cross-track error performance due toGPS-derived tracking errors (Frew et al., 2004; Niculescu2001). Hence, nonlinear controllers based on minimizationof heading or heading error rate were proposed (Frewet al., 2004; Niculescu 2001). Alternative line tracking strate-gies related to proportional navigation–type guidance lawssuch as biased proportional navigation (Holt & Beard, 2010)and precision guidance (Bruggemann et al., 2010) havealso been proposed. Vector field and Lyapunov-based ap-proaches have also been proposed for more general pathtracking problems with UAVs (Nelson, et al., 2006; Ren &Beard 2004).

The past control literature acknowledges the problemof maintaining features of interest within the sensor field ofview; however, the issues associated with tracking powerline corridors and their impact on LiDAR swath width have

not been well studied. The swath width of a LiDAR is theacross-track distance that is mapped while on a survey line(Costa, Battista, & Pittman, 2009). Swath width depends onLiDAR scan angle and also aircraft height above terrain.In tracking power line corridors, it is desired to capture acorridor region. The swath width is an important param-eter in survey applications as the swath needs to be wideenough such that the LiDAR scanning beam covers thecomplete width of the corridor region to be surveyed, yetnarrow enough such that the point cloud density is denseenough to capture specific (and possibly thin) features of in-terest such as individual power lines. Because LiDAR datafiles are large, narrow swath width also has advantages interms of reducing the amount of data stored during a sur-vey flight. Integrated system design, automatic guidance,and maneuver and control solutions for an aircraft track-ing linear infrastructure has already been investigated byBruggemann et al. (2010). However, the impacts on LiDARswath width when tracking power line corridors and theability to conduct automatic turns around power line cor-ridors were not investigated in this earlier work. For thispurpose an evolved prototype of the architecture presentedin Bruggemann et al. (2010) called the Power line Track-ing Automatic Guidance System (PTAGS) was developedas described next.

2.2.2. Power Line Tracking Automatic Guidance System

PTAGS provides improved lateral control for tracking cor-ridors and automatic real-time dynamic turning for flyingbetween corridor segments. A diagram of the system ar-chitecture is given in Figure 3. PTAGS takes waypoints de-scribing the power line corridors as input from a flight andwaypoint management system and commands a standardaircraft autopilot, while considering both translational androtational motion with respect to the corridor. A key fea-ture of this approach is the feedback loop from aircraftperformance (aircraft position and orientation with respectto the power line corridor as measured by onboard navi-gation systems) to PTAGS, which then utilizes new auto-mated guidance, maneuver, and control algorithms to com-mand the aircraft autopilot. It also includes automatic de-cision making, which enables the autopilot to plan and flyturns that incorporate available knowledge at the time suchas LiDAR swath width, aircraft airspeed, and roll angle.Dynamic turn replanning (“go-around”) capability in the

Flight and Waypoint Management System

PTAGS Aircraft

Autopilot Aircraft

Performance

Closed feedback loop

Figure 3. Automatic control of inspection aircraft with PTAGS.



RVWP and Guidance

Maneuver Selection

Autopilot and Dynamics

cφ ca

Feedback of aircraft position and orientation

RVRR WP andGuidance

ManeuverSelection

cφcaPTAGS

Figure 4. Control system architecture of PTAGS.

case of missed data capture or unacceptable cross-track orroll angle tolerances is also implemented.

The control system architecture of PTAGS is given be-low (Figure 4) as already presented in Bruggemann et al.(2010). The key functions of PTAGS are the Receding Vir-tual Waypoint (RVWP), Guidance and Maneuever Selec-tion, which sends roll guidance commands to the aircraftautopilot. The architecture is decomposed and approxi-mately sorted according to the time scales of the dynam-ics handled within each loop. The outer loop (guidance)corresponds to slower features of the dynamics (transla-tional motion), and the inner loop (autopilot) controls thefaster dynamics (rotational motion). This decomposition al-lows the inner autopilot loop to be handled by an already-certified commercial autopilot; the use of certified autopi-lots simplifies the process for certification of the overalldeveloped inspection solution. In fact, we have safely con-ducted flight experimentation with a Cessna 172 aircraft forthe past 2 years, with appropriate approval by the regula-tors. This decomposition also allows platform-specific is-sues to be contained within one loop such that the designof the outer loops is less dependent on platform specifics,allowing reuse on different aircraft platforms.

The trajectory planning and guidance functions are in-dicated by the RVWP and Guidance block in Figure 4. TheRVWP is specified as a point on the infrastructure at somelook-ahead distance ahead of the aircraft’s current location,and the point moves along the infrastructure as the aircraftmoves along. A feature of this approach is the avoidanceof discontinuities in the planned flight path (due to cornerswhere power line segments join), which results in “cuttingacross corners” behavior. When using a RVWP approachto describe the reference trajectory, the lateral stability ofthe platform depends on characteristics of the autopilot re-sponse and the look-ahead distance selected.

The guidance function in PTAGS utilizes a precisionguidance law that commands a lateral acceleration ac tominimize the track error between desired flight trajectoryand current aircraft location. The lateral acceleration com-mands are calculated from current aircraft location, currentRVWP location, and the heading direction of the power lineinfrastructure to be tracked. Further details are published inBruggemann et al. (2010).

The maneuver selection block includes switching logicbetween control for “on survey” and “off survey” tasks. On

survey is when the LiDAR system will be switched on andcollecting data, over the power line corridor infrastructure.In this case the power line corridor itself describes straight-line piecewise linear paths to be tracked and defines thereference trajectory for the RVWP. But there are also “offsurvey” periods in which the aircraft is conducting maneu-vers such as steady turns to align itself in preparation forthe next “on survey” inspection leg. In “off survey” peri-ods a curved reference trajectory for conducting a turningmaneuver is constructed based on Dubins paths (Chitsaz& LaValle, 2007), and the RVWP moves along this curvedpath with tracking provided by the precision guidance law,to achieve an automatic turning control. Roll commands φc

are then sent to the autopilot, which have been determinedfrom lateral acceleration commands ac using a simple kine-matic bank turn model as described in Bruggemann et al.(2010). Studies of the performance of PTAGS in “off survey”mode tracking a curved reference trajectory and impacts ofthe “on survey” line tracking mode on LiDAR swath widthwill be presented in Section 3.

2.3. Multisource Feature Fusion for ImprovedVegetation Classification

LiDAR is an effective sensor for 3D information acquisi-tion and has great potential to assist vegetation manage-ment in power line corridors. However, due to the varia-tions of point density and lack of spectral information, itis often hard to achieve robust tree detection results fromonly LiDAR data. Distinct spectral signatures in red andNIR bands have been successfully used to discriminate veg-etation and nonvegetation (Li et al., 2009). However, grassand low vegetation are difficult to discriminate from treesas they present very similar colors and even textures. A bet-ter solution is to combine multispectral images and LiDARdata to improve tree detection and segmentation.

As discussed in Section 2.1, the species information isvaluable to model the growth of vegetation and discrim-inate desirable and undesirable tree species. Tree speciesclassification from remote sensing data has been inten-sively studied in forest management (Breidenbach, Næsset,Lien, Gobakken, & Solberg, 2010; Holmgren, Persson, &Soderman, 2008; Sugumaran, Pavuluri & Zerr, 2003). How-ever, successful classification is mostly seen at forest standlevel or individual trees with different genera (e.g., conif-erous or deciduous species) because they are obviously dif-ferent in visual appearance. Detailed individual tree speciesare often hard to discriminate due to their similarity in vi-sual features, even for human expert interpretation fromimagery or field survey. We borrow the idea commonlyused in human face recognition and try to improve treespecies classification by selecting more discriminative vi-sual feature descriptors. A feature fusion method is devel-oped by combining color and texture feature descriptorsusing kernel PCA and maximum likelihood–based intrin-sic dimensionality estimation.



2.3.1. LiDAR and Multispectral Data Fusion for ImprovedTree Crown Segmentation

Detection of trees has been intensively studied previously,particularly within the application of remote sensing of for-est environments. Although similar in concept, the environ-ment in a power line corridor is more complex because thebackground is cluttered with shadows, bare soil, shrubs,and grass, all presenting irregularities that need to be han-dled by the detection algorithm. Vegetation has a distinc-tive spectral signature, characterized by a low reflectancein the visible part of the solar spectrum and a high re-flectance in the NIR region. Therefore, NIR information iswidely used in the remote sensing community for the de-tection and classification of vegetation. Combining LiDARelevation data can further improve tree detection by re-moving low-growing grass and shrubs. For classificationof tree species, object-based methods are preferred as theyare straightforward and have been shown to obtain higherclassification accuracy in high-resolution image classifica-tion (Blaschke, 2010). To conduct an object-based classifi-cation, accurate individual tree segmentation is required.Subsequent to this, a range of classification algorithms canbe used in the object-feature space.

As discussed in Section 2.1.2, 2D image–based treecrown segmentation algorithms often meet difficulties indiscriminating grass, shrubs, and trees because these typesof vegetation are often very similar in both color and tex-ture. The 3D nature of LiDAR data makes them especiallywell suited to this situation. However, the success and qual-ity of the results depend on the point density of LiDAR dataas well as the size, shape, and distribution of trees. Multi-spectral imagery provides spectral and texture informationthat is complementary to LiDAR information. Therefore,combining LiDAR data and multispectral imagery seemsto be a promising way to improve individual tree crowndetection and delineation. In this study, a region level fu-sion method is developed to combine multispectral imageand LiDAR data for individual tree crown detection anddelineation.

The first step toward fusing LiDAR and multispectralimagery is referencing. This step is also known as sensoralignment or registration and establishes a common ref-erence frame for different sensor data. If the two sensorsare mounted on the same aerial platform, then the navi-gation system [GPS/inertial measurement unit (IMU)] pro-vides position and attitude data for both the aerial cameraand the LiDAR system. Because the GPS/IMU unit and thetwo sensors are physically separated, the success of sensoralignment relies on how well the relative position and atti-tude of the various system components can be determined.The multispectral imagery and LiDAR data used in ourexperiments have already been georeferenced by the com-mercial data provider, which simplifies the sensor fusionprocess.

Assuming that sensor alignment has been completedwith sufficient accuracy, LiDAR data can then be consid-ered as an additional image layer of the multispectral im-agery. After ground filtering, object points are obtained torefine the tree crown segmentation. The fusion process isdescribed in Figure 5. First, LiDAR point cloud data andgeoreferenced multispectral imagery are processed sepa-rately. On one side, an initial segmentation is conductedin spectral feature space using the algorithm developed inLi et al. (2009). After that, regions in the initial vegetationsegmentation map are labeled for the following fusion pro-cess. On the other side, a ground filtering algorithm usingstatistical analysis is conducted to separate terrain and ob-ject points (Liu et al. (2009). Then the point clouds are con-verted to a 2.5D height image. The pixel size of the heightimage used in this paper is 15 cm, which is the same withmultispectral image pixel. The lowest Z coordinate (height)within the bounds of a pixel is chosen as the height of thatpixel. The 2.5D depth image is then integrated with the la-beled vegetation segmentation map. A simple thresholdingprocess is used in order to remove grass and low vege-tation. The region mean height histogram is calculated tovisualize the height difference among regions. The meanheight of a region that contains grass and low vegetation

Figure 5. Framework of LiDAR and georeferenced multispectral imagery fusion for individual tree crown segmentation.



points will be much lower than that of a region that con-tains only trees. Finally, a watershed-based segmentation(Bleau & Leon, 2000) is employed to further decompose thetree clusters into individual trees.

2.3.2. Color and Texture Feature Fusion for ImprovedTree Species Classification

After individual trees have been segmented, a number oflocal feature descriptors can be used to represent each treecrown. The use of appropriate features to characterize anoutput class or object is fundamental for any classifica-tion problem. There is no generically best feature for im-age classification. The selection of an appropriate featuredescriptor must reflect a specific classification task in handand usually needs to be obtained through experimentalevaluation. We have conducted a feature evaluation (Li,Hayward, Zhang, Jin, & Walker, 2010) and developed anew rotation and scale-invariant, spectral-texture featuredescriptor (Li et al., 2011). The experimental results pub-lished in our previous work demonstrated that tree speciesclassification performance can be considerably improvedthrough careful design of feature descriptors. In this paper,we further evaluate the fusion of multiple color and texturefeatures based on a kernel PCA–based method.

Color and texture are two fundamental features in de-scribing an image, but prior research has generally focusedon extracting color and texture feature as separate entitiesrather than a unified image descriptor (Whelan & Ghita,2009). The idea of using both color and texture informationhas strong links with human perception, and these linksmotivate an investigation into how to effectively fuse colorand texture as a unified descriptor to improve the discrim-ination over viewing color and texture features indepen-dently. Although the motivation of using color and textureinformation jointly in object-based image classification isclear, how best to combine color and texture in a unifiedobject descriptor is still an open issue. Huang, Zhang, andLi (2008) proposed a multiscale spectral and spatial featurefusion method based on wavelet transform and evaluated

in very high-resolution satellite image classification. Zhang,Zhao, Huang, and Li (2008) extracted texture features usingmultichannel Gabor filters, and Markov random fields in-tegrated the two features using a neighborhood oscillatingtabu search approach for high-resolution image classifica-tion. However, these methods extract features from fixedwindow size and do not consider all pixels within an objectas a whole. Moreover, heavy computational burden is in-duced by combining multiple features (through “the curseof dimensionality”), and these burdens may limit the prac-tical performance of the classifier.

It is often difficult to classify objects using a singlefeature descriptor. Therefore, feature-level fusion plays animportant role when multiple features are used in theprocess of object classification. The advantages of featurefusion are as follows: (1) the most discriminatory informa-tion from original multiple feature sets can be derived bythe fusion process; (2) the noisy information can be elim-inated from the correlation between different feature sets.In other words, feature fusion is capable of deriving andgaining the most effective and fewest-dimensional featurevectors that benefit the final classification (Yang, Yang,Zhang, & Lu, 2003). The feature fusion framework is il-lustrated in Figure 6. After object segmentation, color andtexture features are extracted from image objects. Differentfeature vectors are normalized and then serially integrated.After that, kernel PCA is used to globally extract the non-linear features from the integrated feature sets as well asto reduce the dimensionality. The intrinsic dimensionalityof the serial fused features is estimated using a maximumlikelihood method to select the target dimensionality fromthe kernel PCA fused feature. Finally, features selected fromkernel PCA are used as the input to classifiers for furtheranalysis.

We first use a serial fusion strategy by simply combin-ing different feature vectors into one set of feature union-vector. Different features vectors are combined into oneset of feature union-vector. As the features are differenton the value scope, they are initialized into range [−1,1]by Gaussian criterion. Consider an n-dimensional feature

Figure 6. Framework of object-level color and texture feature fusion.



vector F = fij , where fij is the j th feature component inFi . Assuming that f ij is a Gaussian sequence, we com-pute the mean mj and the standard deviation δj . Featuref ij is normalized by f ′

ij = ( f ij − mj )/δj . Suppose that

α and β are two feature vectors that are extracted fromthe same image-object. The integrated feature union-vectoris defined by γ = (α/β). Obviously, if feature vector α ism-dimensional and β is n-dimensional, then the dimensionof feature vector γ is m + n.

Traditional linear feature selection and extractionmethods such as PCA are conducted in the original inputspace and thus cannot properly handle nonlinear relation-ships in the data (Cao, Shen, Sun, Yang, & Chen, 2007).For example, the principal components of features may notbe linearly related to the input variables and the featuresof different categories may not be simply separated by ahyperplane. To solve this problem, kernel methods can beintroduced to map original data to a kernel space usinga mapping function. Kernel PCA is one of these kernelmethods that reformulate traditional linear PCA in a high-dimensional space using a kernel function. Given M inputvectors xp , kernel PCA first maps the original input vectorsxp into a high-dimensional feature space φ(xp). Perform-ing PCA in the high-dimensional feature space can obtainhigh-order statistics of the input variables, which is also theinitial motivation of kernel PCA (Xie & Lam, 2006). In PCA,the principal component of xp is the product of xp and theeigenvectors of the covariance matrix of M input vectors.However, it is difficult to directly compute both the covari-ance matrix of the high-dimensional feature space φ(xp)and its corresponding eigenvectors and eigenvalues in thehigh-dimensional feature space. Therefore, kernel tricks areemployed to avoid this difficulty and the principal eigen-vectors are computed from the kernel matrix, rather thanthe covariance matrix. More details of kernel PCA are in-troduced in the Appendix.

Using kernel PCA to find the underlying structure andthe correlations of multiple feature sets has important ben-efits. First, the most discriminatory information can be de-rived and redundant information can be eliminated fromthe fusion process. Second, the dimension of feature setscan be reduced and thus the computational cost of the sub-sequent classification stage is reduced. However, it can stillbe difficult to find a suitable criterion for selecting optimalfeatures using kernel PCA.

Up to this point we have assumed that the target di-mensionality of the low-dimensional feature representationwas known and specified by the user. In practice, the opti-mal dimensionality needs to be estimated automatically. Apossible solution is to estimate the intrinsic dimensionalityof the high-dimensional feature set and then use this esti-mate as the target dimensionality. Intrinsic dimensionalityis the minimum number of variables that is necessary inorder to represent all the information in a data set. In thispaper a maximum likelihood estimator (MLE) (Levina &

Bickel, 2004) is employed to estimate the intrinsic dimen-sionality. MLE is a local intrinsic dimensionality estimatorthat is based on the observation that the intrinsic dimen-sionality of the data manifold around one data point can beestimated by measuring the number of data points coveredby a hypersphere with a growing radius. MLE considersthe data points in the hypersphere as a Poisson process, inwhich the estimated intrinsic dimensionality d around datapoint xi in given k nearest neighbors is given by

dk(xi ) =⎡⎣ 1

k − 1

k−1∑j=1

logT k(xi )T j (xi )

⎤⎦

−1

, (1)

where Tk(xi ) represents the radius of the smallest hyper-sphere with center xi that covers k neighboring data points.

Different numbers of neighboring data points can betreated as the different scales. It is clear from Eq. (1) that thecalculation of intrinsic dimension d depends on scale pa-rameter k. In this paper, the intrinsic dimension is obtainedby averaging d over a scale range [k1, k2] using the follow-ing equations:

dk = 1M

M∑i=1

dk(xi ), (2)

d =∑k2

j=k1dk

k2 − k1 + 1, (3)

where d is the number of input vectors, dk is the esti-mated dimension at scale k, and d is the final estimateddimensionality.

3. AIRBORNE PLATFORM CONTROL EXPERIMENTSAND RESULTS

3.1. Study of Swath Width Variation underAutopilot Control

Simulated and experimental flight test results of new guid-ance and control techniques for inspection of linear in-frastructure that demonstrated improved cross-track andheading error performance in line tracking were previouslypublished in Bruggemann et al. (2010). Here we present fur-ther results that illustrate the effect of aircraft position andorientation on swath coverage and corridor tracking per-formance under autopilot control. An experimental flighttest was made of a Cessna 172 fitted with PTAGS track-ing a 10-km section of power line corridor at Kingaroy,Queensland, Australia. PTAGS commanded the aircraft au-topilot for lateral control while the pilot maintained verticalcontrol with average speed of 46 m/s and average altitudeof 457 m above ground level. Aircraft position and orienta-tion were recorded via a survey-grade dual frequency GPS–INS system for later evaluation of the flight performance.Wind conditions included a 15-k southwesterly windand low turbulence. The next two subsections study the



Aircraft in straight and level flight

Altitude Above Ground Level (AGL) changes with terrain height

Terrain height variation

Figure 7. Changes in terrain height cause changes in AGL,which will impact swath width.

impact of terrain, altitude, and aircraft, bank angle on Li-DAR swath width.

3.1.1. Impact of Terrain Variation on Swath Width

In this section, the impact of terrain variation on swathwidth is studied on the basis of aircraft flight data, includ-ing measured aircraft position and attitude, collected dur-ing a power line inspection flight. A LiDAR sensor was notinstalled during this flight, but a fictitious LiDAR swathwidth (corresponding to a LiDAR swath pattern coveringa 45-deg field of view) can be calculated from simple geom-etry. The situation is illustrated by Figures 7 and 8, show-ing that the swath width for a lower altitude above ground(AGL) is smaller than the swath width for higher AGL (un-der level terrain assumption).

Figure 9. Aircraft steady climb (calculated), required toachieve a mean altitude of 457 m (1,500 ft) above terrain.

To isolate the impact of the terrain variations, the mea-sured aircraft altitude will be ignored in the followingswath calculations and only aircraft attitude and positioninformation will be used (a mean AGL of 457 m or 1,500 ftwill be assumed). A digital terrain elevation model was ob-tained for the flight test area, and then a fictitious steady-rate climb flight profile was calculated that corresponds toa mean altitude of 457 m (1,500 ft) above the terrain model,as shown in Figure 9. This fictitious vertical profile was as-sumed so that variations in estimated swath would be dueonly to terrain variation and roll motion (and not due tofeatures of the experiment aircraft’s altitude dynamics).Note that in practice there would be some vertical error in

Swath width smaller than powerline corridor due to reduced height above terrain

Swath width covering entire corridor width

Altitude Above Ground Level (AGL)

Aircraft

Ground

Powerline corridor width

FOV

Figure 8. The effect of AGL on swath width. The swath width for a lower AGL is smaller than the swath width for a higher AGL.



Figure 10. Impact of terrain variation on swath width show-ing a difference of up to 86 m (at around 580 s) as compared toflat terrain.

tracking the vertical flight profile under autopilot or man-ual pilot control, which would introduce additional varia-tion in swath width.

Using simple trigonometry, the swath width was thenestimated on the basis of the terrain profile model, the as-sumed vertical flight profile, the measured aircraft position,and attitude, and the width estimates for two different ver-tical flight profiles are shown in Figure 10. Because verti-cal effects have been removed, the variability of the esti-mated swath width is purely due to aircraft roll motionrecorded from the flight test. For comparison purposes, theestimated swath width variation under the alternative as-sumption of a flat terrain profile (that is, constant altitudeflight above a flat terrain) is also shown in Figure 10. Duringthe flight experiment there was an average 26-m and maxi-mum 86-m swath width difference between considering theterrain variations and assuming flat terrain. The noticeablegreater variability in swath width in the presence of terrainvariation suggests that terrain variation impacts the swathwidth more significantly than aircraft roll motion. Thishighlights inspection challenges in terms of not only lat-eral but also vertical automatic control of the aircraft, andthese challenges will impact LiDAR sensor considerations.However, this does not mean that roll motion should be dis-regarded, as will be seen in the next section.

3.1.2. Impact of Aircraft Bank Angle on Swath Width

Results examining swath width variation due to bank angleare now presented (while tracking a 15-km-long section ofpower line corridor using an aircraft under autopilot con-trol). As illustrated by Figure 11, a nonzero bank angle canchange the geometry, causing the swath width to change,

Swath width at non-zero bank angle does not cover entire corridor width.

Aircraft at non-zero bank angle φ

Ground

Powerline corridor width

φ

FOV

Figure 11. Impact of aircraft bank angle on swath width (withflat ground assumption).

but more significantly a nonzero bank angle might causethe swath width to not cover the entire corridor width.

Figure 12 shows the swath variation (estimated swathincluding terrain variation, as explained in the precedingsection) due to bank angle for a Cessna 172 tracking powerlines under autopilot control using the precision guidanceand control algorithms presented in Bruggemann et al.(2010). A desirable corridor width of 200 m was assumedand is displayed on the figure as the red lines.

The PTAGS successfully maintained swath coverageover 98% of the corridor region. However, one part of thecorridor was missed due to excessive roll angle, as indi-cated by A on the figure. Note that during the flight ex-periment the aircraft roll angle did not exceed 15 deg (dueto the rate-1 turn limitation of the autopilot).

From these results it is evident that the aircraft bank(which is necessary to change the heading of the aircraftand maintain lateral track over the corridor) has the un-desirable but unavoidable effect of moving the swath re-gion away from the corridor to be inspected. Thus, autopi-lot guidance and control algorithms for aerial inspectionmust make trade-offs between accurate position track overthe corridor and constraining the roll so that the completecorridor region is measured. Both desired roll changes (toeffect a heading change) and undesired roll changes (suchas due to wind gusts) present a challenge for automaticcontrol of the airborne platform for tracking power linecorridors. This problem is magnified in smaller aircraft orUAVs that are more susceptible to wind gusts. Althoughroll stabilized sensor systems exist, stabilization of the sen-sor head does not solve all issues associated with excessiveroll angle. For example, excessive aircraft roll might causeloss of lock of GPS satellites (GPS is typically integratedin LiDAR survey systems to provide the aircraft’s absoluteposition). Also, the amount of roll stabilization offered by



Figure 12. Tracking a 15-km-long power line corridor underautopilot control with PTAGS. Black, Power line corridor cen-ter, red, 200-m-wide power line corridor; green, swath width(accounting for terrain variation and aircraft bank angle); blue,aircraft ground track. The results demonstrate successful cov-erage of the corridor under autopilot control for most of thecorridor, with occasional miss of corridor (as indicated by Aand B.

commercially available LiDAR aerial survey systems is of-ten limited to small angles due to the additional expensein terms of captured data quantity or quality. It is there-fore desirable to mitigate unwanted roll motion via designof better control of the aircraft platform. Better aircraft rollbehavior might be achieved via maneuver selection strate-gies and the use of constrained bank or skid turns as pro-posed in Bruggemann et al. (2010) and Mills, Ford, & Mejias(2011).

A feature of the specialized power line tracking andguidance algorithms is that the aircraft ground track cutsacross the power line corners (observe the behavior of theblue line of Figure 12). A small corner of corridor that wasmissed due to track error (introduced by the “cutting acrossthe corner” feature of the guidance algorithm) is indicatedon Figure 12 by B. This highlights an issue in tracking short

line segments with sharp changes in line direction—a de-cision is required on whether to proceed to “cut across thecorner” or command a complete turn such that the com-plete segment is captured. This is essentially a trade-off be-tween flight efficiency and certainty of data capture. That is,a complete turn could be conducted at every change in cor-ridor direction to help ensure that corners are not missed,but this would be inefficient in terms of flying time anddistance. On the other hand, a “cutting across corner” ap-proach at every change in corridor direction may result inmissing parts of the corridor as seen in the figure. This de-cision could be made in flight based on the known angu-lar change in corridor direction with a predictive swathcoverage model or made in preflight planning. The au-tomatic turn performance of PTAGS in conducting com-plete turns around corners will be presented in the nextsection.

3.2. Demonstration of Automatic Turn Capabilityaround Corridor Corners

When tracking power line corridors, a key requirement isthat when the aircraft turns around between two line seg-ments, the aircraft can successfully maintain lateral stabil-ity while returning on track toward the next line segmentwithout missing the corner of the corridor where the twosegments meet. This section presents results that illustratethe automatic turn capability of PTAGS around corners.A test case consisting of three line segments connected by90-deg changes in direction was constructed, as shown bythe solid lines in Figure 13. These line segments were flownby the Cessna under autopilot control using PTAGS, whichautomatically calculated and flew turns between each seg-ment. If PTAGS detected that it could not get back ontrack to the next line segment without missing the corri-dor, PTAGS automatically triggered a “go-around again”maneuver.

In Figure 13 the ground track of the aircraft is shownby the red lines (the trajectories shown correspond to thepower lines being inspected twice). The aircraft success-fully executed 90-deg cornering maneuvers, without thesensor swath width missing the corners. During one ofthe cornering maneuvers, the aircraft missed the start ofthe power line corridor at the top corner due to windgusts. However, the PTAGS automatically triggered a “go-around again” maneuver to ensure that inspection is com-plete (hence the three turns shown at the top cornercompared to two turns shown at the bottom corner). Thisautomated “go-around again” maneuver demonstrates thedynamic replanning capability of the system under autopi-lot control. Finally, some lateral instability can be seen onstraight segments, which may indicate that some algorithmtuning or refinement is required.



Figure 13. Automatic turning around 90-degree corners withPTAGS, demonstrating successful tracking of lines after turnsand triggering of a “go-around again” maneuver. Black lines,line segments; red lines, aircraft ground track.

4. AUTOMATED DATA PROCESSING EXPERIMENTSAND RESULTS

4.1. Data Collection

In this study, both aerial remote measurements and groundsurvey data were collected. The first series of flights (con-ducted by a local company) collected data from a high-resolution digital four-band multispectral camera (Duncan-Tech MS-4100) with a DGPS/INS mounted in the cargoarea of a Piper Cub. In these flights, multispectral datawas captured over four spectral bands: NIR (800–966 nm),

Figure 15. Experiment test site.

red (670–840 nm), green (540–640 nm), and blue (460–545 nm) while traveling at approximately 34 m/s (65 kn)at 350 m AGL (multi spectral images were captured at ap-proximately 15-cm spatial resolution). The second seriesof flights (conducted by a different local company) col-lected data from an airborne LiDAR scanner mounted ontoa Cessna aircraft. In these flights, the LiDAR data were col-lected with a scan angle of ±30 deg with an average samplerate of nine points per square meter. Figure 14 shows a pic-ture of our image data collection systems.

The ground survey was conducted in a 1.5-km corri-dor in the towns of Murgon and Wondai in Central EastAustralia where the above-identified multispectral imageand LiDAR data were collected. The ground truth data ofvegetation species were obtained with domain experts’ par-ticipation. Figure 15 shows a mosaic of the test area gener-ated from aerial images acquired from the trial. It shouldbe noted that classifying all types of species in power linecorridors requires significantly more resources than are cur-rently available; however, classifying species in a given testarea as a proof of concept is possible. In this research, wefocus on three dominant species in our test field: Eucalyptustereticornis, Eucalyptus melanophloia, and Corymbia tesselaris.We abbreviate the species names to Euc-Ter, Euc-Mel, andCor-Tes.

Figure 14. Data collection systems.



Figure 16. A pair of the collected multispectral image and LiDAR point cloud data.

4.2. LiDAR and Image Data Fusion

Figure 16 shows color infrared image and LiDAR pointcloud data in urban areas. Figure 17 shows the fusion pro-cess and the result using the image and LiDAR data. Asshown in Figure 17(a), an initial segmentation was con-ducted on the CIR image without any postprocessing. Theinitial segmentation detects trees as well as other vegetationsegments (e.g. grass). Figure 17(b) shows the 2.5D depthimage of LiDAR object points after ground filtering. Eachconnected region in the initial segmentation map was la-beled, showing different colors in Figure 17(c). The 2.5Ddepth image is then integrated with the labeled vegeta-tion segments map. A simple thresholding process is usedin order to remove grass and low vegetation. An overlayof the LiDAR points after the fusion process on the ini-tial segmentation map is shown in Figure 17(d). From thisfigure, we can see that the low-mean-height regions rep-resenting grass and other low vegetation were separated.Figure 17(e) shows the tree segments after the fusion pro-cess. Afterward, a watershed algorithm was applied [Fig-ure 17(e)] to decompose the tree clusters into individualtree crowns. As can be seen from the final segmentationresult, low-vegetation regions have been successfully re-moved. However, a critical limitation of this fusion processis that it depends on the high point density of LiDAR data.For a small-sized tree crown and low-point-density LiDARdata, no points or only a few points hit the tree, whichmay cause the tree to be removed due to low region meanheight. A series of 13 pairs of multispectral imagery and Li-DAR data were selected for processing, with a total numberof 183 trees. Frames were removed from the full sequenceof images to minimize image overlap (overlap might seesome trees processed more than once). Furthermore, the de-veloped algorithm was applied only to those areas of theimage that contained the power line corridor. The devel-oped tree crown segmentation algorithm achieved a detec-tion rate of 97.27% and a segmentation accuracy of 84.7%.

4.3. Color and Texture Feature Fusion

To select the most appropriate features for tree speciesclassification, we evaluated a number of state-of-the-artlocal feature descriptors and machine learning classifiers(Li et al., 2011; Li, Hayward, Zhang, et al. 2010). The eval-uation results demonstrated that the classification successvaries between different feature descriptors and classifiers.Overall, the rotational and scale invariant spectral-texturefeature developed in Li et al. (2011) showed the best over-all classification accuracy, and the support vector machine(SVM) was suggested as it generally obtains robust clas-sification performance. In this paper, we extend the samestrategy to evaluate the performance of the feature fusionmethod with a SVM classifier. The V-fold cross-validationtechnique was employed in the experiment, and 10 fold-ers were selected for the cross validation. The data set ispartitioned into 10 groups, which is done using stratifi-cation methods so that the distributions of categories ofthe target variable are approximately the same in the par-titioned groups. Then 9 of the 10 partitions are collectedinto a pseudo-learning data set, and a classification modelis built using this pseudo-learning data set. The rest 10%(1 of 10 partitions) of the data that were held back and usedfor testing the built model, and the classification error forthe data was computed. After that, a different set of ninepartitions is collected for training and the rest (10%) is usedfor testing. This process is repeated 10 times, so that everyrow has been used for both training and testing. The clas-sification accuracies of the 10 testing data sets are averagedto obtain the overall classification accuracy.

Two classic color and texture features, color his-togram and local binary pattern (LBP) (Ojala, Pietikainen, &Maenpaa, 2002), are tested in the experiment. The overallclassification accuracies of the fused color-texture featurewere compared with single color and texture feature vec-tors and serial integrated feature vector through the sameclassifier. For comparison purpose, another widely used



Figure 17. Fusion of LiDAR and multispectral imagery for individual tree crown segmentation.

nonlinear feature selection technique, generalized discrim-inant analysis (GDA), is also evaluated in the experiment.GDA is also known as kernel linear discriminant analy-sis (kernel LDA); it is the reformulation of LDA in thehigh-dimensional space constructed using a kernel func-

tion (Baudat & Anouar, 2000). In this experiment, a Gaus-sian kernel function is used to construct GDA for the fusionof color-texture features.

From the experiment, the overall classification accura-cies of color histogram and LBP texture features are 76.03%



Table I. The classification results of single and fused color and texture features.

Hist RGBNIR LBP Serial fusion KPCA-40 GDA-40

Overall accuracy (%) 76.03 71.07 83.47 95.04 91.73Analysis time (s) 79.46 246.29 305.83 13.63 4.63

Figure 18. Classification accuracies of the fused features at different dimensions.

(a) Tree crown detection and segmentation

(b) Classification map (species are represented using different colors)

Figure 19. An example of individual tree segmentation and species classification.



and 71.07%, respectively. The serial integration of thesetwo features shows better performance over single fea-ture, with an overall accuracy of 83.47%. To evaluate theperformance of fused feature using kernel PCA, we use astep-by-step model justification method (Song & Tao, 2010).We justify the dimensionality from 2 to 8 with steps of2, and from 10 to 100 with steps of 10, for the fused fea-ture vectors. Figure 18 shows the classification accuracycurve at different dimensions. As we can see from the fig-ure, the kernel PCA fused feature performs much betterthan single feature and serial integrated feature. However,the kernel PCA fused feature is still based on the assump-tion that the user can specify a good target dimensional-ity. The estimation of intrinsic dimensionality using MLEis employed as the automatic selection of optimal num-ber of dimensions. From our experiment, the intrinsic di-mensionality of the integrated LBP and color histogramfeature is 39.3016, which conforms to the result from Fig-ure 18, where the best accuracy (95.04%) is obtained atdimension 40.

In the experiment, the computational costs of the clas-sifiers using different feature vectors are also compared.The analysis time is recorded under a desktop PC config-uration of core duo 2.66-GHz CPU and 2 GB of memory.Table I summarizes the overall accuracies and analysis timeusing different feature sets. The optimal dimension of ker-nel PCA and GDA fused feature is 40, which is derivedfrom the estimation of the intrinsic dimensionality. Fromthe results, we can see that the analysis time varies a lotfor different feature sets. High dimensionality of the origi-nal color and texture features and the serial fused featurecause high computational costs, whereas using a nonlin-ear fusion method such as kernel PCA and GDA can notonly improve the classification accuracy but also signifi-cantly reduce the dimensionality and the computationalcosts.

From the experimental results, it is clear that fusionof color and texture features provides improved discrim-inative power over using them independently. Moreover,the proposed nonlinear feature fusion strategy using kernelPCA has shown great improvement over the serial fusionstrategy, not only on reducing the dimensionality and com-putational cost, but also on removing noisy informationand improving the discriminative power. The proposedfeature fusion strategy can be extended to combine anyother feature vectors if they are considered to have somecomplementary information.

Figure 19 shows an example of individual tree crownsegmentation and classification map. It should be notedthat trees can often appear differently in different seasonsand even the same tree species may vary due to healthstatus. The available data did not provide the opportu-nity to classify the same vegetation under different condi-tions. Therefore, more variables may need to be consideredto model a specific tree species more accurately in futurework.

5. CONCLUSION AND FUTURE WORK

This paper comprehensively investigated the use of aerialremote sensing techniques for power line corridor monitor-ing and vegetation management. A technology overviewand a series of experiments and results are presented.The major technological contributions and the lessons welearned are summarized as follows:

• Automatic Control of Airborne PlatformCurrently, standard industry practice is to fly power

line corridors under manual piloted control. This modeof flying over extensive power line infrastructure on aroutine basis is seen to be dangerous and tedious, jus-tifying the development of autopilot technologies thatmay assist a human pilot and eventually lead to com-plete autonomous UAV solutions for more efficient andsafer operations.

The Power Line Tracking Automatic Guidance Sys-tem (PTAGS) developed at the Australian Research Cen-tre for Aerospace Automation (ARCAA) specifically fortracking power line corridors attempts to address someof the horizontal control problems and showed satisfac-tory performance in tracking power line corridors. Ex-perimental results demonstrated the impact of roll an-gle and terrain variation on swath width, which must beconsidered for an aircraft under autopilot control that istracking power line corridors. The key challenges lie inthe lateral and vertical automatic control of the aircraftas consideration of both translational and rotational mo-tion is required. Further work is required to addressthese challenges further and extend the approach to alsoconsider vertical motion.

• Reliable and Automated Data ProcessingThe information derived from multisensor data and

various modeling approaches provides the opportunityto increase the reliability of the information extractionfor robust operational performance and decision makingin corridor monitoring (e.g., improved classification, in-creased confidence, and reduced ambiguity). By fusingLiDAR with multispectral imagery and also color withtexture features, we were able to significantly improvetree segmentation and species classification. The nextstage is to apply the same fusion strategy to combinegeometric features derived from LiDAR data and thespectral-texture features derived from multispectral im-agery. Effective fusion of multisensor, multiresolution,multitemporal, and multiplatform image data, togetherwith geospatial data and GIS, represents the future solu-tion for smart power line corridor monitoring.

The presented methods for object recognition (i.e.,power lines and trees) are designed for offline use. Dataare collected, stored in a repository, and analyzed at somelater time. Real-time processing of data would be beneficialif useful information could be extracted that assists in thedecisions made by the aircraft control system. For example,



real-time power line detection can be used for the purposesof active aircraft guidance in situations in which there is noor limited prior knowledge about the network (e.g., GPSlocations of a power line are not accurately known). Alter-natively, the real-time identification of regions outside theimmediate vicinity of the power line where vegetation issparse could be used to reduce the resolution of data stored.Questions concerning the algorithms and computing archi-tectures that are best suited to increase the autonomy of aUAV capturing significant amounts of data need to be ad-dressed in future work.

6. APPENDIX: KERNEL PRINCIPAL COMPONENTANALYSIS (PCA)

Assuming that the kernel matrix is centered, i.e.,∑Mp=1,q=1 K (xp, xq ) = 0, introducing the kernel matrix

K makes the mapping implicit without manipulating high-dimensional space φ(xp) explicitly in terms of Mercer’scondition. Each element of K is an inner product in thehigh-dimensional space [i.e., K (xp, xq ) = φ(xp) · φ(xq )].Assuming that C is the covariance matrix of φ(xp), μi andβi are the ith eigenvalue and eigenvector of C, respec-tively; λi and αi are the ith eigenvalue and eigenvectorof the kernel matrix the K . The relationships betweeneigenvectors and eigenvalues of C and K are

μi = λi

M, (A.1)

βi =M∑

j=1

αji × φ(xj ), (A.2)

where αji is the j th element of αi ( j = 1, . . . , M).

To obtain low-dimensional feature representation, thedata are projected onto the eigenvectors of the covariancematrix C. The result of low-dimensional data representa-tion Y is obtained by computing the principal eigenvectorsof components of xp in the space φ(xp):

yip =

M∑j=1

αji × K (xj , xp), (A.3)

where yjp is the ith element of yp and αi is the normalized

αi (αi = αi/√

λi ).The mapping performed by kernel PCA relies on the

choice of the kernel function. In this paper, Gaussian kernel,which is widely used in many applications, is employed.The kernel function is defined as

K(xi, xj ) = exp

(−

∥∥xi − xj

∥∥2

2δ2

), (A.4)

where δ is the shape parameter.

ACKNOWLEDGMENTS

This work was conducted within the CRC for Spa-tial Information, established and supported underthe Australian Government’s Cooperative ResearchCenters Programme and sees collaboration betweenthe Queensland University of Technology (QUT),the Australian Research Centre for Aerospace Au-tomation (ARCAA), and ROAMES Ergon EnergyAustralia. The authors gratefully acknowledge thecontributions of Professor Rodney Walker, Dr. Ross Hay-ward, and Mr. Ryan Fechney. The contributions of BredJeffers from Greening Australia and David Wood fromErgon Energy for their assistance in the field survey arealso acknowledged.

REFERENCES

Baudat, G., & Anouar, F. (2000). Generalized discriminant anal-ysis using a kernel approach. Neural Computation, 12(10),2385–2404.

Beltrame, A. M. K., Jardini, M. G. M., Jacbsen, R. M., &Quintanilha, J. A. (2007, July). Vegetation identifica-tion and classification in the domain limits of power-lines in Brazilian Amazon forest. In IEEE InternationalGeoscience and Remote Sensing Symposium, Barcelona,Spain (pp. 2314–2317).

Berni, J. A. J., Zarco-Tejada, P. J., Suarez, L., & Fereres, E. (2009).Thermal and narrowband multispectral remote sensingfor vegetation monitoring from an unmanned aerial vehi-cle. IEEE Transactions on Geoscience and Remote Sensing,47(3), 722–738.

Blaschke, T. (2010). Object-based image analysis for remotesensing. ISPRS Journal of Photogrammetry & RemoteSensing, 65(1), 2–16.

Bleau, A. e., & Leon, L. J. (2000). Watershed-based segmenta-tion and region merging. Computer Vision and Image Un-derstanding, 77(3), 317–370.

Breidenbach, J., Næsset, E., Lien, V., Gobakken, T., &Solberg, S. (2010). Prediction of species specific forest in-ventory attributes using a nonparametric semi-individualtree crown approach based on fused airborne laser scan-ning and multispectral data. Remote Sensing of Environ-ment, 114, 911–924.

Bruggemann, T. S., Ford, J. J., & Walker, R. A. (2010). Control ofaircraft for inspection of linear infrastructure. IEEE Trans-actions on Control Systems Technology, pp(99), 1–1.

Cao, B., Shen, D., Sun, J.-T., Yang, Q., & Chen, Z. (2007, June).Feature selection in kernel space. In Proceedings of the24th International Conference on Machine Learning, Cor-vallis, OR.

Chaput, L. J. (2008, February). Understanding LiDAR data—How utilities can get the maximum benefits from 3D mod-elling. In International LiDAR Mapping Forum. Denver,CO.

Chitsaz, H., & LaValle, S. (2007, December). Time-optimalpaths for a dubins airplane. In IEEE Conference on De-cision and Control. New Orleans, LA.



Costa, B. M., Battista, T. A., & Pittman, S. J. (2009). Compara-tive evaluation of airborne LiDAR and ship-based multi-beam SoNAR bathymetry and intensity for mapping coralreef ecosystems. Remote Sensing of Environment, 113(5),1082–1100.

Deng, Y., & Manjunath, B. S. (2001). Unsupervised segmenta-tion of color-texture regions in images and video. IEEETransactions on Pattern Analysis and Machine Intelli-gence, 23(8), 800–810.

Egbert, J., & Beard, R. W. (2007, July). Low altitude roadfollowing constraints using strap-down EO cameras onminiature aerial vehicles. In American Control Confer-ence, New York.

Erikson, M. (2003). Segmentation of individual tree crownsin colour aerial photographs using region growing sup-ported by fuzzy rules. Canadian Journal of Forest Re-search, 33(8), 1557–1563.

Erikson, M., & Olofsson, K. (2005). Comparison of three indi-vidual tree crown detection methods. Machine Vision andApplications, 16(4), 258–265.

Frew, E., McGee, T., Kim, Z., Xiao, X., Jackson, S., Morimoto,M., Rathinam, S., Padial, J., & Sengupta, R. (2004, March).Vision-based road-following using a small autonomousaircraft. In IEEE Aerospace Conference.

Gurtner, A., Greer, D. G., Glassock, R. R., Mejias, L., Walker,R., & Boles, W. W. (2009). Investigation of fish-eye lensesfor small-UAV aerial photography. IEEE Transactions onGeoscience and Remote Sensing, 47(3), 709–721.

Holmgren, J., Persson, A., & Soderman, U. (2008). Species iden-tification of individual trees by combining high resolu-tion LiDAR data with multi-spectral images. InternationalJournal of Remote Sensing, 29(5), 1537–1552.

Holt, R. S., & Beard, R. W. (2010). Vision-based road-followingusing proportional navigation. Journal of Intelligent andRobotic Systems, 57(1–4), 193–216.

Huang, X., Zhang, L., & Li, P. (2008). A multiscale feature fu-sion approach for classification of very high resolutionsatellite imagery based on wavelet transform. Interna-tional Journal of Remote Sensing, 29(20), 5923–5941.

Ituen, I., & G. Sohn (2010). The way forward: Advances inmaintaining right-of-way of transmission lines. Geomat-ica, 64(4), 451–462.

Jones, D., Golightly, I., Roberts, J., Usher, K., & Earp, G. (2005,November). Power line inspection–A UAV concept. InThe IEE Forum on Autonomous Systems, London, UnitedKingdom.

Jwa, Y., Sohn, G., & Kim, H. B. (2009, September). Automatic3D power line resconstruction using airborne lidar data.In ISPRS Laserscanning 2009. Paris, France.

Kobayashi, Y., Karady, G. G., Heydt, G. T., & Olsen, R. G. (2009).The utilization of satellite images to identify trees endan-gering transimission lines. IEEE Transactions on PowerDelivery, 24(3), 1703–1709.

Kwak, D.-A., Lee, W.-K., Lee, J.-H., Biging, G. S., & Gong,P. (2007). Detection of individual trees and estimation oftree height using LiDAR data. Journal of Forest Research,12(6), 425–434.

Leckie, D., Gougeon, F., Hill, D., Quinn, R., Armstrong, L.,& Shreenan, R. (2003). Combined high-density lidar andmultispectral imagery for individual tree crown analysis.Canadian Journal of Remote Sensing, 29(5), 633–649.

Lefsky, M. A., & Cohen, W. B. (2003). Selection of remotelysensed data. In M. A. Wulder and S. E. Franklin (Eds.), Re-mote sensing of forest environments: Concepts and casestudies (pp. 13–49). Dordrecht, The Netherlands: KluwerAcademic Publishers.

Levina, E., & Bickel, P. J. (2004). Maximum likelihood estima-tion of intrinsic dimension. Advances in neural informa-tion processing systems (vol. 17). Cambridge, MA: MITPress.

Li, Z., Hayward, R., Walker, R., & Liu, Y. (2011). A biologicallyinspired object spectral-texture descriptor and its applica-tion to vegetation classification in power line corridors.IEEE Geoscience and Remote Sensing Letters, 8(4), 631–635.

Li, Z., Hayward, R., Zhang, J., Jin, H., & Walker, R. (2010).Evaluation of spectral and texture features for object-based vegetation species classification using support vec-tor machines. International Archives of the Photogram-metry, Remote Sensing and Spatial Information Sciences,XXXVIII(Part 7A).

Li, Z., Hayward, R., Zhang, J., & Liu, Y. (2008, December).Individual tree crown delineation techniques for vegeta-tion management in power line corridor. In Digital ImageComputing: Techniques and Applications (DICTA), Can-berra, Australia. IEEE Computer Society.

Li, Z., Hayward, R., Zhang, J., Liu, Y., & Walker, R. (2009,November). Towards automatic tree crown detection anddelineation in spectral feature space using PCNN andmorphological reconstruction. In IEEE International Con-ference on Image Processing, Cairo, Egypt.

Li, Z., Liu, Y., Walker, R., Hayward, R., & Zhang, J. (2010).Towards automatic power line detection for a UAVsurveillance system using pulse coupled neural filter andan improved Hough transform. Machine Vision and Ap-plications, 21, 677–686.

Liu, Y., Li, Z., Hayward, R., Walker, R., & Jin, H. (2009, Decem-ber). Classification of airborne lidar intensity data usingstatistical analysis and Hough transform with applicationto power line corridors. In International Conference onDigital Image Computing: Techniques and Applications.Melbourne, Australia. IEEE.

Lu, M. L., & Kieloch, Z. (2008). Accuracy of transmission linemodeling based on aerial LiDAR survey. IEEE Transac-tions on Power Delivery, 23(3), 1655–1663.

Mallinis, G., Koutsias, N., Tsakiri-Strati, M., & Karteris, M.(2008). Object-based classification using Quickbird im-agery for delineating forest vegetation polygons in aMediterranean test site. ISPRS Journal of Photogramme-try & Remote Sensing, 63, 237–250.

Mills, S., Ford, J. J., & Mejias, L. (2011). Vision based control forfixed wing UAVs inspecting locally linear infrastructureusing skid-to-turn maneuvers. Journal of Intelligent andRobotic Systems, 61(1), 29–42.



Mills, S. J., Gerardo, M. P., Li, Z., Cai, J., Hayward, R., Mejias,L., & Walker, R. (2010). Evaluation of aerial remote sens-ing techniques for vegetation management in power linecorridors. IEEE Transactions on Geoscience and RemoteSensing, 48(9), 3379–3390.

Nelson, D. R., Barber, D. B., McLain, T. W., & Beard, R. W. (2006,June). Vector field path following for small unmanned airvehicles. In American Control Conference, Minneapolis,MN.

Niculescu, M. (2001, January). Lateral track control law foraerosonde UAV. In The 39th AIAA Aerospace SciencesMeeting and Exhibit, Reno, NV.

Ojala, T., Pietikainen, M., & Maenpaa, T. (2002). Multiresolu-tion grey-scale and rotation invariant texture classificationwith local binary patterns. IEEE Transactions on PatternAnalysis and Machine Intelligence, 24(7), 971–987.

Pouliot, D. A., King, D. J., Bell, F. W., & Pitt, D. G. (2002).Automated tree crown detection and delineation in high-resolution digital camera imagery of coniferous forest re-generation. Remote Sensing of Environment, 82(2–3), 322–334.

Pouliot, D. A., King, D. J., & Pitt, D. G. (2005). Developmentand evaluation of an automated tree detection-delineationalgorithm for monitoring regeneration coniferous forests.Canadian Journal of Forest Research, 35(10), 2332–2345.

Rathinam, S., Kim, Z., Soghikian, A., & Sengupta, R. (2005,December). Vision based following of locally linear struc-tures using an unmanned aerial vehicle. In The 44th IEEEConference on Decision and Control, and the EuropeanControl Conference, Seville, Spain.

Rautiainen, M. (2005). The spectral signature of coniferousforests: The role of stand structure and leaf area index.Helsinki, Finland: Faculty of Agriculture and Forestry.University of Helsinki.

Ren, W., & Beard, R. W. (2004). Trajectory tracking for un-manned air vehicles with velocity and heading rate con-straints. IEEE Transactions on Control Systems Technol-ogy, 12(5), 706–716.

Song, D., & Tao, D. (2010). Biologically inspired feature man-ifold for scene classification. IEEE Transactions on ImageProcessing, 19(1), 174–184.

Sugumaran, R., Pavuluri, M. K., & Zerr, D. (2003). The use ofhigh-resolution imagery for identification of urban climaxforest species using traditional and rule-based classifica-tion approach. IEEE Transactions on Geoscience and Re-mote Sensing, 41(9), 1933–1939.

Sun, C., Jones, R., Wu, H. T. X., Cheong, K., Beare, R., Buckley,M., & Berman, M. (2006). Measuring the distance of vege-tation from powerlines using stereo vision ISPRS Journalof Photogrammetry & Remote Sensing, 60(4), 269–283.

Whelan, P. F., & Ghita, O. (2009). Color texture analysis. InM. Mirmehdi, X. Xie, and J. Suri (Eds). Handbook oftexture analysis (pp. 129–164). London: Imperial CollegePress.

Xie, X., & Lam, K.-M. (2006). Gabor-based kernel PCA withdoubly nonlinear mapping for face recognition with a sin-gle face image. IEEE Transactions on Image Processing,15(9), 2481–2492.

Yan, G., Li, C., Zhou, G., Zhang, W., & Li, X. (2007). Automaticextraction of power lines from aerial images. IEEE Geo-science and Remote Sensing Letters, 4(3), 387–391.

Yang, J., Yang, J.-y., Zhang, D., & Lu, J.-f. (2003). Feature fu-sion: Parallel strategy vs. serial strategy. Pattern Recogni-tion, 36, 1369–1381.

Zhang, L., Zhao, Y., Huang, B., & Li, P. (2008). Texture fea-ture fusion with neighborhood oscillating tabu search forhigh resolution image. Photogrammetric Engineering &Remote Sensing, 74(12), 1585–1596.


toward automated power line corridor monitoring

Documents

corridor segments

significant pressure

acrosspower lines

continuous electricity

proposed aircraft control

tracking corridors

ergon energy

aerospace automation