QUALITY ASSESSMENT FOR LIDAR POINT CLOUD REGISTRATION USING IN-SITU CONJUGATE FEATURES
- 1. IGARSS, 24-29 July 2011, Vancouver, Canada (Session FR2.T03) Quality Assessment for LIDAR Point Cloud Registration using In-Situ Conjugate FeaturesJen-Yu Han1, Hui-Ping Tserng1, Chih-Ting Lin21 Department of Civil Engineering, National Taiwan University2 Graduate Institute of Electronics Engineering, National Taiwan University
- 2. NTUCvE Surveying Engineering GroupOutlineIntroduction
- 3. Using In-Situ Conjugate Features
- 7. ConclusionIntroductionLight Detection and Ranging (LIDAR) is capable of acquiring 3D spatial information in a fast and automatic manner.
- 8. Can be equipped on platforms of various kinds (air-borne, mobile, and terrestrial).
- 9. Usually requires multiple scans in order to construct a complete and accurate 3D model.Reason 1: Incompleteness due to obstructionsReason 2: Error magnification due to projective geometry
- 10. Introduction (cont’d)Incompleteness due to obstructionsMany obstructions could occur when the LIDAR point cloud is collected from a single station.Only partial information is acquired for the 3D object.
- 11. Introduction (cont’d)Error magnification due to projective geometryPoint coordinates are based on range and angular measurements both of which contain errors.As a result, the quality will become lower for outer regions.
- 12. Introduction (cont’d)Registration of LIDAR datasets from multiple stations Each dataset is defined in an arbitrary local reference frame.A 3D similarity transformation model is usually postulated to relate the datasets defined in different reference frames. 12212s: scaleR: rotation matrixt: translation vector1Station 1 Station 2
- 13. Using In-Situ FeaturesObtaining the transformation parametersClassic approach: point-based least-squares approach
- 14. Find (>=3) conjugate points in two LIDAR datasets
- 15. Perform least-squares parameter estimationsRequires extra effort to set up identifiable targets (e.g. control spheres or reflective sticks) or perform feature extractions.Requires a set of good initial values and iterative computations to obtain reliable parameter estimates.
- 16. Using In-Situ FeaturesObtaining the transformation parametersProposed approach: using directly in-situ features Extended feature types Definite features Points: vectors between points Lines: directional vectors Planar patches: normal vectorsIndefinite features Groups of points: eigenvectors of the tensor field constructed by a group of point.With these extended feature types, it becomes possible to use the geometric components that are already inherent in the scanned object.
- 17. Using In-Situ FeaturesIn-situ features usable for LIDAR dataset registrationsHighway surfaces Bridge pillarsSlope surfaces and edges Structure edges and railsNo need to set up control targets reduce the cost for field work.
- 18. Weighted NISLT ApproachOnce feature correspondence is established, the transformation parameters are estimated by the weighted NISLT (Non-Iterative Solutions for Linear Transformations) technique:Scale parameterwhere dxij and dx’ij are coordinate differences (vectors) in the original and transformed systems, is the weight matrix, lkis a kx1 unity vector.
- 19. Weighted NISLT ApproachRotational parameterswhere ΔX and ΔX’ are the matrices by stacking all the normalized row vectors in the original and transformed systems. Translational parameters
- 20. Weighted NISLT ApproachCharacteristics of weighted NISLT approach - Closed-form solution, requires no initialvalues nor iterative computations highly efficient compared to LSQ-based approaches. - Weighted parameter estimation model uncertainties of input observables can be realistically taken into consideration. - Accepts input observables of different kinds (e.g. vectors between points, directional vectors of linear features, normal vectors of planar features, and eigenvectors of groups of points) make possible a direct use of various in-situ geometric features.
- 21. Quality AssessmentClassical point-based approach: Registration quality is typically evaluated by the post-fit residuals for point coordinates after applying the estimated parameters. : post-fit residual vector of point in : number of conjugate pointsThis index gives a vague interpretation on the obtained result since it represents only the positional agreement between two datasets geometrical similarity is not considered!!
- 22. Quality AssessmentProposed approach: Here features of various kinds are used for a registration. The quality is then evaluated based on the following two indexes:Absolute Consistency (qa) Relative Similarity (qr)Positional alignment Geometric similarity: post-fit residual vector of conjugate point i or the vector between point i ‘s projected points on two conjugate features. : the angle between two conjugate vectors (directional vectors, normal vectors, or eigenvectors) after the registration. : the numbers of conjugate points and conjugate vectors
- 23. (a) (b)(c) (d)Quality AssessmentInterpretation of a registration solution: Moderate qa, good qr. Moderate qa and qr. Poor qa, good qr. Poor qa and qr.The quality of a registration solution can be explicitly defined by the proposed two indexes qa and qr.
- 24. S2S1Numerical ValidationData collection: A case study was performed for a 250m-long reinforced concrete (RC) bridge in Taipei City.Two LIDAR stations (S1, S2) were set up about 80m away from the bridge.
- 25. Numerical ValidationData collection (cont’d): LIDAR point cloud was collected at each station using a Trimble® GS200 Terrestrial Laser Scanner.Resolution for the scanned points of the bridge was roughly between 0.02m ~ 0.04m.No control sphere or reflective stick was set up in the scanned area. TrimbleGS200 Laser Scanner - Range: 2m~200m - Accuracy: range = 6 mm @ 100 mangular = 6 mm @ 100 m - Max. Density: 3mm@100m
- 26. Numerical ValidationCollected datasets and in-situ features used for registrationTwo sets of LIDAR point clouds were collected at the two stations.Since no control point was available, in-situ features were selected from the datasets and used for a registration.Two pillars, a rail and a beam surface were used as conjugate features. Station 1 Station 2
- 27. Numerical ValidationNISLT registrationThe eigenvectors of conjugate features were used as observables while solving for the transformation parameters based on the proposed weighted NISLT approach.Station 1 Station 2
- 28. Numerical ValidationRegistration results (integrated point clouds)Shown in true colorsShown in blue for points collected at station 1 and in red for points collected at station 2
- 29. Numerical ValidationRegistration results (integrated point clouds)S2S1Integrated
- 30. Numerical ValidationRegistration results (quality assessment)
- 31. Absolute consistency (qa) = 3.81cm.
- 32. Relative similarity (qr) = 1.864e-4 .
- 33. qr is equivalent to a 3.73cm positional distortion for an object of size 200m. Equally accurate in terms of positional agreement and geometric similarity. Both values are within a reasonable range considering the 2cm~4cm resolution of the original LIDAR datasets the registration quality is mostly dependent on the point resolution in this case.
- 34. ConclusionThe proposed approach increases the number of usable features for a registration solution the cost for LIDAR field work can be significantly reduced.
- 35. The weighted NISLT enables an efficient parameter estimation when in-situ hybrid conjugate features are used.
- 36. The two quality indexes (absolute consistency and relative similarity) give a complete and explicit quality indication for a registration solution.
- 37. An automatic approach for selecting qualified in-situ features should be developed in the future.Thanks for your attention For more information, please contact: Jen-Yu Han, Ph.D. Department of Civil Engineering, National Taiwan University Email: jyhan@ntu.edu.tw Phone: +886-2-33664347 Website: http://homepage.ntu.edu.tw/~jenyuhan