Transcript
VALIDATION AND TUNNING OF DENSE STEREO-VISION SYSTEMS USING HI-RESOLUTION 3D REFERENCE MODELS
E. REMETEAN (CNES) S. MAS (MAGELLIUM) - JB GINESTET (MAGELLIUM) - L.RASTEL (CNES)
ASTRA - 14.04.2011
CONTENT
Validation methodology Hardware & software tools Preliminary results Study status Credits
ASTRA 2011
2
Dense stereo-vision system validation To build a navigation map and compute a safe and effective trajectory, the Autonomous Navigation software needs a reliable knowledge of the rover surroundings
Study goals 3D reconstruction accuracy Robustness wrt the scene content & lighting Impact of the stereo-vision system parameters Optimal parameter set definition for a given robotic mission
Methodology Comparison of the computed disparity maps to dense reference maps acquired with low parallax
ASTRA 2011
3
Acquisition Mechanical Ground Support Equipment Acquisition of stereo images & reference models with reduced parallax Composed of a stereo-bench, a Laser Scanner, translation linear stages
FARO Photon 20 Laser Scanner Up to 30 million 3D points / s-b f.o.v. Accuracy ≈ ±2mm Measurement range up to 20m Translation stages for parallax minimization
ASTRA 2011
4
Acquisition MGSE calibration Estimation of transfer matrix between the Laser Scanner and stereo-bench reference frames Parallax minimisation
Stereo Bench (SB)
Laser Scanner (LS)
Calibration performed after any MGSE displacement on the Mars yard
Laser Tracker (LT)
Laser Tracker used to measure Position of the landmark balls Position & attitude of the stereo-bench Accuracy better than 0.1mm
Landmark Balls
Useful area
ASTRA 2011
5
Acquisition campaigns Both indoors and outdoors (scene content & lighting conditions variation) Several exposure times for every scene (robustness studies) The content of the scene is changed rather than moving the MGSE (to avoid MGSE calibrations)
ASTRA 2011
6
Laser Scanner data filtering 3D-Filter software was developed for: Interest zone selection (corresponding to stereo-bench field of view) Points clouds filtering for measurement artefacts & outliers removal
ASTRA 2011
7
Data exploitation: Perception Workshop Comparison of real disparity (physical stereo-bench) to virtual disparity (virtual stereobench looking at the 3D model measured by the LS) Real disparity computation parameters can be modified from a control panel The filtered Laser Scanner points cloud is meshed and the disparity is computed using the virtual stereo-bench physical parameters (adjustable) Initial virtual stereo-bench position & attitude obtained from MGSE calibration step ASTRA 2011
8
Data exploitation: PW – Stereo-benchmarking Similarity scores Classical window-based scores (SAD, SSD, ZNCC) 3D-distance score (a little pessimistic) Virtual 3D cloud
Real 3D cloud Left optical centre
3Ddist = | DepthReal – DepthVirtual |
Virtual stereo-bench position & attitude optimisation Stereo base length optimisation → indirect stereo base length measurement method ASTRA 2011
9
Data exploitation: PW – 3D viewer 3D viewer allows to display in 3D Real & Virtual points clouds computed from disparities Mismatches between the clouds (local similarity error)
ASTRA 2011
10
Preliminary results (1/3) Accuracy (3Ddist) 100mm stereo base CCD 4.65µm pixels Full resolution images
Scenes
Mean Error (mm)
Std dev (mm)
Indoor
5.05
0.64
Outdoor
13.27
1.55
7x7 correlation window Virtual SB attitude & base optimisation to measure intrinsic performance
☺ Mean Error < Autonomous Navigation DEM cell size (40mm) Impact of image resolution Image subsampling
“Pixel size” (µm)
Mean error (mm)
Mean accuracy degradation
Number of pixels to process
1/1
4.65
13.27
0%
100%
1/2
9.30
16.60
25.09%
25%
1/4
18.60
20.45
54.11%
6.25%
ASTRA 2011
11
Preliminary results (2/3) Impact of stereo-correlation window size Correlation window size
Mean error (mm)
Mean accuracy gain wrt 7x7
Estimated complexity wrt 7x7
9x9
12.45
6.18%
+65%
7x7
13.27
0%
0%
5x5
15.00
-13.04%
-51%
Robustness to exposure time L\R
5 ms
48 ms
81 ms
87 ms
135 ms
170 ms
5 ms
27.24
-
-
-
-
-
55 ms
-
88.50
66.87
57.10
15.33
7.27
90 ms
-
75.03
89.68
81.58
47.92
28.70
100 ms
-
62.31
82.68
90.09
77.61
55.69
150 ms
-
19.64
57.60
82.59
90.18
86.70
200 ms
-
8.10
32.99
55.30
87.42
90.19
ASTRA 2011
Correlation ratio (%)
12
Preliminary results (3/3) Fast multi-resolution stereo-correlation algorithm L\R
5 ms
48 ms
81 ms
87 ms
135 ms
170 ms
5 ms
81.90
-
-
-
-
-
55 ms
-
92.98
91.11
92.49
86.82
82.71
90 ms
-
92.74
93.08
92.48
91.14
88.96
100 ms
-
93.11
92.22
93.11
92.79
91.89
150 ms
-
91.97
92.11
92.94
93.19
92.98
200 ms
-
90.33
91.26
92.23
92.90
93.15
L\R
5 ms
48 ms
81 ms
87 ms
135 ms
170 ms
5 ms
+24.5
-
-
-
-
-
55 ms
-
-7.6
-7.4
-10.1
-0.8
+2.1
90 ms
-
-9.3
-8.7
-7.2
-12.7
-14.9
100 ms
-
-14.7
-7.7
-9.4
-11.7
-13.6
150 ms
-
-13.5
-13.0
-11.8
-9.1
-9.2
200 ms
-
-17.7
-18.0
-17.6
-9.1
-9.2
ASTRA 2011
Correlation ratio (%)
Mean error (% wrt mono-resolution algorithm)
13
Study status Today Validation methodology adapted for dense stereo-vision systems Indirect stereo base estimation method Accuracy of the studied stereo-vision system is compatible with AN requirements Good robustness to image exposure conditions Fast multi-resolution algorithm will become the new CNES baseline
Further work Full-resolution outdoor acquisition campaigns Performances with stereo-bench flight model demonstrator “Tough” textures campaigns Stereo base length impacts Security margins definition for Autonomous Navigation
ASTRA 2011
14
Credits : CNES sub-contractors involved Stereo-benches Flight model demonstrator: CSEM / MCSE Ground models: COMAT Aerospace, AR2P
Perception Workshop Magellium CS-SI
3D-Filter CS-SI
Validation studies & MGSE realisation Magellium Sud-Rectif ASTRA 2011
15
Thank you for your attention!
[email protected]
ASTRA 2011
16
