WE2.TO9.2.pptx
- 1. ADAPTIVE MORPHOLOGICAL FILTERING FOR DEM GENERATIONKyoHyouk Kim and Jie ShanGeomatics EngineeringSchool of Civil EngineeringPurdue University July 27th 2011IGARSS 2011 Annual Conference
- 4. INTRODUCTIONLiDAR (Light Detect And Ranging)One of remote sensing technologies providing point clouds with highly accurate collections of (X,Y,Z) at observed points.Vertical accuracy : 6 to 30 cmHorizontal accuracy : 10 to 46 cmUsed as one of primary data sources to applications:DEM (Digital Elevation Model)Forest mapping Building footprints & 3D building reconstruction4/28
- 5. INTRODUCTIONFilteringSeparation between ground points and nonground points.Ground points : Digital Elevation Model (DEM).Non-ground points : Building models, forest structure, canopy mappingIn nearly all applications, filtering should be done first.Various filtering algorithms have been proposed.Slope-based filtersLinear prediction filtersClustering (or segmentation) filtersMorphological filters5/28
- 7. OUTLINEIntroductionPrevious WorksGeneral ruleMorphological filtering algorithmKnown issues & weaknessesProposed ApproachTestsSummary7/28
- 8. PREVIOUS WORKSGeneral ruleWhat is ground surface ?Ground surfaces are relatively smooth & continuousLocally spanned through the lowest parts.EX: discontinuous ground surface and negative blunders.All filtering algorithms try toIdentify points on the continuous and smooth ground surfaceRemove points lying on non-ground objectsMeasures of discontinuityAbrupt slope or elevation difference8/28
- 9. PREVIOUS WORKS Morphological filteringCommonly used in the image processing fields to :Remove noiseEnhance imagesExtract featuresPublished papers for LiDAR filteringKilian, Halla, and Englich (1996), Kilian et al. (1996), Zhang et al. (2003), Zaksek and Pfeifer (2006), Zhang and Cui (2007), Chen et al. (2007,2009). 9/28
- 10. PREVIOUS WORKS PrinciplesMorphological opening operation : Erosion followed by dilationErosion : Dilation :Objects with different sizeApply different window size If terrain is not flatuse maximum sloped0d10/28
- 11. PREVIOUS WORKSKnown issues & weaknessesParameters have a great impact on the filtering resultSlope : allowed minimum elevation differenceWindow size : objects with various sizesSmall window : Can not remove larger buildingsLarge window : Highly possible to flatten ground surfaceNo optimal set of parameters supporting various types of terrainRequire a priori information about the target area.Require trial-and-error process to find the best set of parameters11/28
- 12. OUTLINEIntroductionPrevious WorksProposed ApproachAdaptive morphological filteringTestsSummary12/28
- 13. PROPOSED APPROACHAdaptive morphological filteringUse of regular grid with grid size g ( average point spacing)Cell with multiple points : point with the lowest elevationEmpty cell : elevation of the nearest pointKeeps the indices of original LiDAR points1D (or 2D) Morphological erosion operation (W = 3).Filtering is applied only to points of discontinuity iteratively.Relevant parameters are adaptively adjusted in each iteration13/28
- 14. PROPOSED APPROACHUse of 2D regular grid14/28
- 15. PROPOSED APPROACHDiscontinuity MeasureIdentify points lying on the edge between ground and objectsUse of nominal value for thresholds (No fixed thresholds)Ex: 50, 75 or 90 percentileThreshold is determined before each iterationResidualHeight differenceSlope15/28
- 16. PROPOSED APPROACHExampleOne LiDAR profileSlopeResidual16/28
- 17. PROPOSED APPROACHWorkflowFor i=1 to # of rows (or # of columns)Determine residuals (ri) of all pointsIfri> N percentile of r (N=50,75,90)If hi - (New hi) > Hminhi = New hiUntil # of filtered points = 0End17/28
- 18. PROPOSED APPROACHExample70th row - 50 percentile - Hmin = 1.0mPurdue campus18/28
- 19. PROPOSED APPROACHRefinementType I (omission) error : remove ground points mistakenlyCaused by (1) negative blunders and (2) discontinuous groundType II (commission) error : classify objects to ground pointsCaused by (1) nonground points with smaller elevation than Hmin19/28
- 20. PROPOSED APPROACHExample ( Type I error)20/28
- 21. PROPOSED APPROACHResolve Type I errorIdentify continuous segments of ground pointsDistance of any non-ground points to the line (Si(n) – Si+1(1))Repeated until no more ground points are restoredd21/28
- 22. OUTLINEIntroductionPrevious WorksProposed ApproachTestsSummary of data setsEvaluationsSummary22/28
- 23. INITIAL RESULTSTest data setsHigh relief (mixed slope)
- 26. Almost flat
- 27. Buildings mixed with trees
- 28. High relief (mixed slope)
- 30. INITIAL RESULTSFiltered ground surface (1)Hmin= 1.5m
- 31. Residual threshold = 50 percentile24/28
- 32. INITIAL RESULTSFiltered ground surface (2)Hmin = 1.0 m
- 33. Residual threshold = 50 percentile25/28
- 34. INITIAL RESULTSFiltered ground surface (3)Hmin = 1.5 m
- 35. Residual threshold = 50 percentile26/28
- 37. SUMMARYAdaptive morphological filteringMinimize the effects of parametersFiltering is applied only to points on the boundary iterativelyResidual is used as the measure id discontinuityResidual threshold is adaptively adjusted with nominal valuePromising results from various terrain conditions, but :Type I error is often significant in case of discontinuous terrainNeed more robust stopping criterion28/28