runway centerline deviation estimation from point · pdf file1 runway centerline deviation...
TRANSCRIPT
1
Runway Centerline Deviation Estimation from Point Clouds using LiDAR imagery
ICRAT 2016Drexel University, Philadelphia, PA ♦ June 20-24, 2016
Seth Young1, Charles Toth2, Zoltan Koppanyi21Department of Civil, Environmental and Geodetic Engineering
The Ohio State UniversityPhone: 614-292-7681
2Satellite Positioning and Inertial Navigation (SPIN) LabCenter for Aviation Studies
E-mail: [email protected]
Outline
Introduction Motivation Remote sensing concept
Laser scanning technologies Sensor selection Initial test results
Tests with taxiing aircraft 1-LiDAR sensor configuration LiDAR point cloud processing
• Methods• Method comparison
Tests with touch&go aircraft 4-LiDAR sensor configuration Results and statistical analysis
Tests with landing aircraft 5-LiDAR sensor configuration Results and statistical analysis
Conclusion/Future work
2
4
Potential
Wingtip clearance,airfield separation
Veer-off models
Tracking take-offs and
landings
• Airport planning: stochastic models.
Source: ACRP Report 51, Risk Assessment Method to Support Modification of Airfield Separation Standards
Light Detection And Ranging
Courtesy of Ayman Habib
INS
Airborne LiDAR/ALS
Mobile LiDAR/MLS
5
Google’s driverless car
6
Concept
LiDAR Sensor
Trajectory
Outline
Introduction Motivation Remote sensing concept
Laser scanning technologies Sensor selection Initial test results
Tests with taxiing aircraft 1-LiDAR sensor configuration LiDAR point cloud processing
• Methods• Method comparison
Tests with touch&go aircraft 4-LiDAR sensor configuration Results and statistical analysis
Tests with landing aircraft 5-LiDAR sensor configuration Results and statistical analysis
Conclusion/Future work
7
8
Sensor Comparative Tests
Laser technology is rapidly advancing Application specific sensors are introduced (UAS, autonomous
driving) Performance evaluation needed
to optimize sensor selection to provide reference for low-end sensors
Profilers/scanners considered:
Bridger’s HRS-3D-1W
UTM-30LX-EW
Ibeo Alasca XT
SICK LMS30206 Velodyne HDL-64E
Velodyne HDL-32E
Velodyne VLP-16: 16 channels 100+ m range 300K points/second Dual-return capability 360° x 30° FOV Small size: 100 mm x 65 mm Weight: 600 grams Low power requirements Easy to install No external rotating parts Relatively inexpensive sensor
9
Single Scanner Test
Site location: OSU Don Scott Airport The sensors attached to a data acquisition platform (GPSvan) Test aircraft: Cessna 172 Scenarios: aircraft passing diagonally and perpendicularly w.r.t
sensors at various velocities (3, 6, 9,12 knots)
10
Test 3: Airplane Georeferencing
11
Data Preprocessing
1. Point cloud filtering: remove no aircraft body points, such as runways, buildings, light fixtures, etc.
2. Filtering by time: select those time intervals when aircraft was in the field of view of the sensor
12
Methods: Challenges
Sparse point cloud
Difficult to model aircraft bodySample raw data
13
Methods under Investigation
1. Center of Gravity (COG) Point cloud center of gravity estimated and tracked
2. ICP-based 2-DoF – 3D ICP (only heading and velocity
estimated)3. Volume minimization (VM) – new approach,
specifically developed for the project Bin-based random optimization (entropy
minimization)
14
Method 1: COG
Calculate the center of gravity (COG) of each frame (one point cloud)
15
Method 2: Iterative Closest Point
Problem: find the rotation matrix and translation vector that minimize the sum of the distance of the closest points between two points clouds
Source: Point Cloud Library, http://pointclouds.org/documentation/tutorials/int
eractive_icp.php
where 𝑀𝑀 is the 4-by-4 transformation matrix,𝑠𝑠𝑖𝑖 = [𝑥𝑥𝑠𝑠,𝑖𝑖 , 𝑦𝑦𝑠𝑠,𝑖𝑖 , 𝑧𝑧𝑠𝑠,𝑖𝑖 , 1] is the 𝑖𝑖th point from
the samples,𝑟𝑟𝑖𝑖 = [𝑥𝑥𝑠𝑠,𝑖𝑖 , 𝑦𝑦𝑠𝑠,𝑖𝑖 , 𝑧𝑧𝑠𝑠,𝑖𝑖 , 1] is the closest 𝑖𝑖thcoordinates on the reference body curve
Solve this minimization problem between the consecutive frames
min𝛼𝛼,𝛽𝛽,𝛾𝛾,Δ𝑥𝑥,Δy,Δz
�𝑖𝑖=1
𝑁𝑁𝑝𝑝
[𝑀𝑀(𝛼𝛼, 𝛽𝛽, 𝛾𝛾, Δ𝑥𝑥, Δy, Δz) ∗ 𝑠𝑠𝑖𝑖 − 𝑟𝑟𝑖𝑖]2𝑀𝑀 =
𝑟𝑟1,1 𝑟𝑟1,2𝑟𝑟2,1 𝑟𝑟2,2
𝑟𝑟1,3 Δ𝑥𝑥𝑟𝑟2,3 Δ𝑦𝑦
𝑟𝑟3,1 𝑟𝑟3,20 0
𝑟𝑟3,3 Δ𝑧𝑧0 1
16
Motion Models
Uniform motion Trajectory is a straight line
𝑥𝑥 = 𝑣𝑣𝑥𝑥𝑡𝑡𝑦𝑦 = 𝑣𝑣𝑦𝑦𝑡𝑡
Curvilinear motion Trajectory is a 2nd order
polynomial
𝑥𝑥 = 𝑣𝑣𝑥𝑥𝑡𝑡 +𝑎𝑎𝑥𝑥
2 𝑡𝑡2
𝑦𝑦 = 𝑣𝑣𝑦𝑦𝑡𝑡 +𝑎𝑎𝑥𝑥
2 𝑡𝑡2
Free motionTrajectory is a higher older
polynomial
𝑥𝑥 𝑡𝑡 = 𝑎𝑎0 + 𝑎𝑎1𝑡𝑡 + 𝑎𝑎2𝑡𝑡2 + ⋯ + 𝑎𝑎𝑛𝑛𝑡𝑡𝑛𝑛,𝑦𝑦 𝑡𝑡 = 𝑏𝑏0 + 𝑏𝑏1𝑡𝑡 + 𝑏𝑏2𝑡𝑡2 + ⋯ + 𝑏𝑏𝑛𝑛𝑡𝑡𝑛𝑛
17
Method 3: VM Reconstruction
Problem: reconstruct aircraft body using the equation of motion at various velocities
18
Method 3: Volume metric
Decimating space into cubes (voxels); optimized to average point cloud density; 0.5x0.5x0.5 m was used
The “goodness” of the reconstruction is measured by the number of cubes occupied by the reconstruction
19
Visual Comparison
20
Statistical Comparison
COG may provide good solution but the full aircraft body has to be captured; difficult to achieve
ICP is able to provide acceptable results only in short ranges and when the motion direction w.r.t sensors is close to ~45°; in other cases, it is likely to fail
VM (volume/entropy minimization) gives the best and most robust solution
Outline
Introduction Motivation Remote sensing concept
Laser scanning technologies Sensor selection Initial test results
Tests with taxiing aircraft 1-LiDAR sensor configuration LiDAR point cloud processing
• Methods• Method comparison
Tests with touch&go aircraft 4-LiDAR sensor configuration Results and statistical analysis
Tests with landing aircraft 5-LiDAR sensor configuration Results and statistical analysis
Conclusion/Future work
21
22
Tests 4 and 5
23
System Components
Sensor station With one vertical and one horizontal sensor
Installation (from top to down) Sensor station, GPS antenna, logging laptops, power
GPS AntennaPrecise timing and station coordinates
24
Sensor Configuration
Logging laptop
Horizontal sensor
Verticalsensor
GPS Antenna
Battery
Sensor interface boxGPS Receiver
25
Network Configuration
26
Sample Sensor Data
Landing aircraft
Horizontal sensor axis
Vertical sensor axis
27
Typical Raw Data
28
Point Clouds
Point clouds acquired by all sensors (top view)
Point clouds acquired by all sensors (rear view)
29
Aircraft Variations
Info Sample 1 Sample 2Aircraft type Cessna LearjetOperation Landing TaxiingNumber of sensors 4 2Observation time [s] 7 6.1Observed length [m] ~100 ~27Number of backscattered points 5,970 13,133
30
Trajectory Estimation
31
Derived Data
Outline
Introduction Motivation Remote sensing concept
Laser scanning technologies Sensor selection Initial test results
Tests with taxiing aircraft 1-LiDAR sensor configuration LiDAR point cloud processing
• Methods• Method comparison
Tests with touch&go aircraft 4-LiDAR sensor configuration Results and statistical analysis
Tests with landing aircraft 5-LiDAR sensor configuration Results and statistical analysis
Conclusion/Future work
32
33
Tests 6
34
Sensor Network Configuration
35
Obtaining Reference: Surveying
Goals: Precise centerline reference Precise global coordinates Aiding sensor calibration
Equipment: GPS static solution for control points Total Station for mass points
36
Sensor Configuration
Station 2(2 horizontal)Station 1
(2 horizontal,1 vertical)
GPS Antenna
Sensor
Interface Box
GPS Receiver
Logging laptops, batteries
CablingUTP & Power
37
Sensor Calibration
Sensor position is known from GPS, but the sensor orientation is unknown
The sensor orientation is estimated during the calibration
Three rotation/attitude angles: roll, pitch, yaw
Very accurate orientation determination is required for precise positioning, as data from various sensors should be integrated
SensorX,Y,Z position is known from GPS
FOV: unknown
FOV: unknown
38
Using Planar Target for Sensor and Intra-sensor Calibration
Sensor station
Calibration boardCaptured by the LiDAR
sensors
4 prisms installed at
cornersBoard corners in global system are measured by Total
Station
39
Sensor Calibration
Station
Surface normal-basedorientation estimation.
Reference normal is calculated from TS measurements.
Minimize the angle difference between the boards’ and the reference norms.
This is an overdetermined homogenous linear equation system that can be solved by SVD.
40
Data Acquisition
Sensor station
Landing airplane
41
Method 4: Cube Trajectories
Initial reconstruction from VM
Point association by decimating the space into cubes
The points with different timestamps inside one cube defines the cube’s trajectory
Calculating the cubes’ velocities based on the points
Filtering the velocity time sequence by a Gaussian filter
Trajectory can be derived by integrating the velocity function
42
Estimated Velocities
43
Results (Video)
44
Estimated Altitudes (34 planes)
45
Estimated Altitudes (34 planes)
Outline
Introduction Motivation Remote sensing concept
Laser scanning technologies Sensor selection Initial test results
Tests with taxiing aircraft 1-LiDAR sensor configuration LiDAR point cloud processing
• Methods• Method comparison
Tests with touch&go aircraft 4-LiDAR sensor configuration Results and statistical analysis
Tests with landing aircraft 5-LiDAR sensor configuration Results and statistical analysis
Conclusion/Future work
46
47
Conclusion
LiDAR technology allows tracking taxiing and landing aircraft Single LiDAR sensor configuration may not provide
acceptable results, depending on the sensor quality (affordability)
A multi-LiDAR prototype system developed The interactive target-based sensor network calibration
process provided good accuracy During a real four-hour long test, the landing and/or
taxiong of 34 aircrafts were recorded and processed From the LiDAR point clouds, 2D and 3D trajectories
were estimated ~10 cm level accuracy was obtained; depending on sensor-
object distance, aircraft and sensor types
48
Future Work
Assess the reliability and accuracy of the system using GPS reference on aircraft (DGPS, cm-level accuracy)
The optimal sensor network formation and configuration have still open questions
Tests with larger aircraft Very little or no effort has been devoted to investigate the use of
complementary sensors, such as cameras and radar/UWB Point cloud processing has advanced and more sophisticated feature
extraction techniques could be exploited for plane type detection
There have been rapid sensor developments, so the affordability of a sensor configuration for an entire runway is becoming a reality
Ford signed an agreement with Velodynetargeting a $500 unit price
49
Thank you for the attention!
Q&A