Permanent Structure Detection in Cluttered Point Clouds from Indoor Mobile Laser Scanners (IMLS)

Transcription

1 Permanent Structure Detection in Cluttered Point Clouds from NCG Symposium October 2016 Promoter: Prof. Dr. Ir. George Vosselman Supervisor: Michael Peter

2 Problem and Motivation: Permanent structure reconstruction, wall detection, room classification Opening detection from cluttered data: door, window Applying current methods on MLS data Topology reconstruction Permanent Structure Detection in Cluttered Point Clouds from Top view of fire brigade building Front view of the same room Top view of one room 2

3 Related Work: Wall detection: wall is a permanent structure dividing 3D space Methods: 2d histogram (Adan & Huber, 2011), normal vector (Sanchez and Zakhor, 2012), cell decomposition (Mura et al., Oesau et al., Xiao and Furukawa, 2014; Ochmann et al., 2016), density histogram (Iro Armeni et al. 2016) Problems: excess wall detected, some walls are missed, clutter close to walls, occlusion Opening detection from cluttered data: ray-tracing (Adan & Huber, 2011) Wall detection and room classification, Mura et al Iro Armeni et al

4 Wall detection using topology relation Opening detection using occlusion test Door detection in voxel space using trajectory Room classification Permanent Structure Detection in Cluttered Point Clouds from Objectives and Workflow: Improving permanent structure 4

5 Segmentation Intersect segments Permanent Structure Detection in Cluttered Point Clouds from Wall detection: Wall detection using topology relation (Oude Elberink, S.J., 2015) Labeling: segment is labeled as wall: iff is connected to the ceiling AND other walls Check the result with room candidates Intersection lines Top view of segmentation wall-segment wall-segment Wall Other labels floor Labeling process Top view of wall detection 5

6 Wall detection result: Permanent Structure Detection in Cluttered Point Clouds from Zeb1 data from Fire brigade building (top view) 6

7 Permanent Structure Detection in Cluttered Point Clouds from Opening detection: Opening detection using occlusion test (Adan & Huber, 2011) Point clouds from Zeb1 and NavVis M3 MLS trajectory as sensor position Zeb1 data from Fire brigade building (top view) 7

8 Permanent Structure Detection in Cluttered Point Clouds from Opening detection: Opening detection using ray-tracing Point clouds and trajectory (top view) Point clouds and trajectory Candidate surface points (front view) 8

9 Permanent Structure Detection in Cluttered Point Clouds from Opening detection : Opening detection using occlusion test Generate planar voxel grid from candidate surface Label voxels as occupied and unoccupied Occupied Unoccupied Surface point cloud (front view) Voxels (front view) 9

10 Permanent Structure Detection in Cluttered Point Clouds from Opening detection: Opening detection using occlusion test A voxel on the wall is opening if there is a point behind the wall, otherwise occluded. occupied Intersected voxel Occupied Unoccupied opening Before occlusion test (front view) Trajectory Occlusion occluded Points behind the surface Occupied Occluded Opening Unoccupied Wall Plane (right side view) After occlusion test 10

11 Door detection: Permanent Structure Detection in Cluttered Point Clouds from Voxelize the point clouds Find open door centers with three rules: 1. A door center is in empty space 2. Above the door center there are points 3. There should be a trajectory close by Extract door borders Extract closed doors For closed doors trajectory goes through a wall Front view: point clouds containing two doors closed door (front view) Trajectory (white) and door centers (red) 11

12 Results: Fire brigade building 3 rd floor 12

13 Room classification and navigable space: Store empty space within a margin of clutter and walls between floor and ceiling Store Navigable space above the floor (right image) Subdivide the empty space with door locations (left image) Top view of empty spaces. Black areas representing walls and clutter. Navigable space just above the floor Yellow circles are location of doors 13

14 Room classification and navigable space: Store empty space within a margin of clutter and walls between floor and ceiling Store Navigable space above the floor (right image) Subdivide the empty space with door locations (left image) Top view of empty spaces. Black areas representing walls and clutter. Navigable space just above the floor Yellow circles are location of doors 14

15 Room classification and navigable space: Store empty space within a margin of clutter and walls between floor and ceiling Store Navigable space above the floor Subdivide the empty space with door locations Top view of empty spaces. Black areas representing walls and clutter. 15

16 Analyzing the methods: Permanent Structure Detection in Cluttered Point Clouds from Wall Detection: relies on the segmentation and connectivity of segments. Opening Detection: relies on the wall detection results, challenge in occluded openings and reflection from glass. Door Detection: relies on the trajectory and input door size parameter. Room Classification: windows and gaps in the data are problematic for space subdivision. Advantages of our method: + Applicable on non-manhattan World + Applicable on non-vertical walls + Scalable to large datasets + Improvable with iterations Disadvantages of our method: - Big gaps in the data challenge topology reconstruction -??? 16

17 Permanent Structure Detection in Cluttered Point Clouds from Conclusion and further work: Changing the order of steps and iteration is expected to improve the results voxelization segmentation door detection wall detection with topology empty space subdivision occlusion test room classification door / windows/gaps 17

18 Future Plan: Evaluate the result of the object detection Improving the topology and geometry of the generated model Design the shape grammar rules Apply the shape grammar on the generated model for large buildings 18

19 Thank You for your Attention Questions? 19