spinning laserchitresh/papers/csiro_talk_chitresh.pdf · references j.l lerma, j.m biosca:...
TRANSCRIPT
Spinning Laser
Chitresh BhushanUndergraduate student, IIT Kharagpur, India
Under supervision of
Mike BosseSenior Research Scientist, QCAT,
CSIRO ICT Center, Brisbane
Objective
To detect and join the edges, in real-time,in 3D point cloud generated using range data from a single laser sensor, which can be used for mapping, navigation etc.
What is spinning laser ?
Simple laser sensor rotating in a plane perpendicular to scan-plane (as shown).
Provides :
Range information Angular Velocity Orientation of sensor
Pre-processing
Correcting data with calibration corrections.
Updating the timestamps to allow some clock-drift.
Validating all scans depending on updated timestamps of neighboring scans.
Filtering & Processing
Phantom discontinuity is removed with range data of neighboring point.
Data is median filtered for 5 points in each scan.
3D Cartesian coordinates, pose, normals & weights are calculated for each point using neighboring points.
Edge Detection I
Zero crossing of 2nd order derivative is used over the range data.
Laplacian of Gaussian (LoG) is used. (sensitive to edges with smoothing out the noise)
5x5 kernel is used with standard deviation of 0.7 .
0.083307 1.018699 2.123167 1.018699 0.0833071.018699 5.519411 0.300248 5.519411 1.0186992.123167 0.300248 -40.254 0.300248 2.1231671.018699 5.519411 0.300248 5.519411 1.0186990.083307 1.018699 2.123167 1.018699 0.083307
Edge Detection II LoG kernel is applied on last 5 scans & all other processing
are done using last 2 scans, of which LoG values are known
False Edges I Theoretical error:
Presence of a zero crossing in between the actual edges.
f(x)=Gaussian of i(x)
C(x)=f '(x)
Ref: Clark, J. J., Authenticating edges produced by zero-crossing algorithms, IEEE Transactions on Pattern Analysis and Machine Intelligence, 11(1):43-57 (1989).
False Edges II Very few points with zero LoG value
-127
0(Blue)
+127
False Edges III Practically too many zero crossing
-127
0(Blue)
+127
Removing False Edges I Using minimum threshold in difference of
LoG value at zero crossing. Removes false edges due to noise.
Points with +ve LoG value
Red: Required edge(nearer points)
Green: False edge
Range
Removing False Edges II
Nearer points are edges
Removing False Edges III Normal condition number: confidence of
normals calculated.
Ratio of two smallest eigen values of scatter matrix.
Lower normal condition number corresponds to a flat region.
Thresholding a minimum value for normal condition number removes false edges due to theoretical error in zero crossing.
Removing False Edges IV
Blue: Low normal cond. No Red: High normal cond. No
Fitting Planes
Joining edges Following the zero crossing & assigning
each edge point on that zero crossing same Edge ID.
Joining edge points with same ID, in correct order.
Edges crossing each others are NOT joined.
Results
Results
Results
Results
Results
Results
Results
Results
Results
Results
Results
Results
Results
Results
Results
Results
Live Demonstration !
Future Work I Direction of normals for edge detection
Edges, which do not have discontinuity in range values, can be detected by considering the direction of normals in neighboring points.
Future Work II Using more information from history
Edges are not joined when any one point in intermediate scan is missed. Looking beyond one scan line may solve the problem.
Future Work III Adaptive thresholds
Range values increases rapidly when scan points are at large distance from sensor (point P3 & P4, above) & it may lead to a false edge at P3. An adaptive (with range) threshold can fix the problem.
Future Work IV Hexagonal neighborhood
8-neighbors 6-neighbors
Acknowledgment
Mike Bosse Paul Flick Robert Zlot Felix Duvallet All QCAT staff
References J.L Lerma, J.M Biosca: Segmentation and filtering of laser scanner data for cultural
heritage. In CIPA, page 896, Torino, Italy, 2005 Qiang Ji, Robert M. Haralick: Quantitative Evaluation of Edge Detectors Using the
Minimum Kernel Variance Criterion. ICIP (2) 1999: 705-70 Yon-Lin Kok, Soheil I.Sayegh, Joo-Heng Hong: An Algorithm to Find Two-
Dimensional Signals with Specified Zero Crossings. IEEE Transactions on acoustics, speech and signal processing, Vol. ASSP-35, No. 1, January 1987: Page 107
Fangwei Zhaol, Christopher J.S. Desilva, Use of the Laplacian of Gaussian operator in prostate ultrasoundimage processing. Proceedings of the 20th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Vol. 20, No 2,1998
Boulaassal, H., Landes T., Grussenmeyer P., Tarsha-Kurdi F., Automatic Segmentation of Building Facades Using Terrestrial Laser Data, IAPRS Volume XXXVI, Part 3 / W52, 2007
Ben Weiss, Shell & Slate Software Corp. Fast Median and Bilateral Filtering, ACM Transactions on Graphics (TOG) 2006, 519 - 526
C. Chu, N. Nandhakumar, and J. K. Aggarwal, "Image segmentation using laser radar data,"Patt. Recogn., vol. 23, no. 6, pp. 569-581, 1990.
Questions ?
Thank You