Rapid Extraction and Updating Road Network from LIDAR Data

Transcription

1 Rapid Extraction and Updating Road Network from LIDAR Data Jiaping Zhao, Suya You, Jing Huang Computer Science Department University of Southern California October, 2011

2 Research Objec+ve Road extrac+on from remotely sensed data is a tradi+onal, but challenging topic Challenge: accuracy, confidence, completeness, and automa+on Comparing with 2D imageries, LIDAR data has 3D informa+on acached Our goal is to explore the advanced airborne LIDAR sensing data for automated road extrac+on and road quality mapping grid- structured urban road network

3 Data prepara+on Oakland LIDAR data Intensity image Depth image

4 Grid- structured road network extrac+on Major Steps 1. Ground Object Separa3on 2. Road Region Detec3on 3. Intersec3on Detec3on 4. Centerline Input: LIDAR data Data Conversion 2D intensity image 2D depth image Hierarchical Morphology Ground objects Elevated objects EM Classifica>on Road candidate image Radius- Rota>ng Sliced candidate image TLS Line Fi9ng Road centerlines 5. Road Network Comple3on Output: Vectorized road network Vote- based Removal & Missing Line Inference

5 1. Ground objects separa+on Goal: separate elevated and ground objects from depth image Method: Hierarchical Morphological Opening Approximate DTM Elevated objects = depth image approximate DTM Ground objects = (Elevated objects) C

6 1. Ground objects separa+on Oakland

7 1. Ground objects separa+on Denver

8 2. Road feature classifica+on Goal: road candidate extrac+on from ground objects mask and LIDAR intensity image Instead of classifying the whole intensity image, we carry out classifica+on on grounded objects only Assume intensity distribu+on of ground objects to be Gaussian Mixture K ( ) = ( ) ( ) = π (, ) k 1 k µ = k Σk p x p z p x z N z Employ Expecta+on- Maximiza+on(EM) algorithm to find the maximum likelihood solu+on for Gaussian Mixture models Each pixel is classified according to their posterior probability

9 2. Road feature classifica+on EM Intensity image Ground objects mask Road candidate image EM

10 3. Road intersec+on detec+on AXer EM classifica+on, ground road candidates (binary image) are separated. From this raster image, we proceed to detect road centerlines. Here, we develop our approach. First we propose a Radius- Rota>ng method to find road intersec+ons, and cut roads apart from these intersec+ons. Now originally con+nuous roads are sliced into disconnected segments. AXerwards, total least squares fi[ng is employed to es+mate centerlines for each segment.

11 3. Road intersec+on detec+on Radius- Rota>ng: Unlike template matching methods, in which various shape models like T, +, L shapes are defined, here we only have to put a radius on the candidate point, rotate it and count the number of road regions it goes across. Radius- Rota>ng method is illustrated in the following figure:

12 3. Road intersec+on detec+on a) 3 branches b) 2 branches a) 3 branches: intersec+on candidate point; b) 2 branches: further check the angle between the two direc+ons.

13 3. Road intersec+on detec+on

14 3. Road intersec+on detec+on

15 4. TLS road centerline fi[ng For each separated segment, we use total least squares(tls) to fit a center line. Given the line equa+on "#+$%+&=0, the criterion of TLS is to minimize the sum of perpendicular distances between points and the line, i.e. we need to minimize: i ( ax + by + c) 2 i i

16 4. TLS road centerline fi[ng Road candidate image TLS line fi[ng

17 5. Road network comple+on AXer TLS line fi[ng, there can be many false posi+ves (non- road centerlines). Before proceeding to missing road inference, we d becer remove these false posi+ves as many as possible. In general, false posi+ves have random direc+ons, while true road centerlines share several main direc+ons. Based on this observa+on, we propose the direc>on- based cumula>ve vo>ng process to pre- remove those centerlines with minor direc+ons (small voted values).

18 5. Road network comple+on Major direc>on: 5 green lines and 5 blue lines, each of them is voted by other 4 with similar direc+on. So each has cumula>ve voted value 4; Minor direc>on: randomly distributed (red, orange and black), there is no vote for them. So each has cumula>ve voted value 0.

19 5. Road network comple+on TLS line fi[ng cumula+ve vo+ng result Minor direc+on: removed

20 5. Road network comple+on AXer preliminary cumula+ve- vo+ng based segments removal, we adopt the following procedure for network inference and valida+on: Geometrical constraints guided gap filling Hypothesis test (back- projec+on valida+on) Geometrical constraints: 1) Collinearity 2) Gap 3) Curvature 4) Length. Hypothesis test: AXer inference of these possible gaps, we back- project them onto the intensity and depth image, and perform the valida+on using intensity and eleva+on within the regions.

21 5. Road network comple+on Based on these 4 measures, we define failing- link- rate by linear combina+on: flr = αa+ β D + µ C + νl the smaller, the more possible a) Collinearity: direc+on difference A1 A2 A = Amax if ( A A < A ) b) Gap: gap length gap D = Gmax gap < G 1 2 max otherwise max otherwise c) Curvature: direc+on variance C = T T + T T d) Length: longer road first L min L = min( L, L )

22 5. Road network comple+on Road network comple+on based on geometrical constraints and back- projec+on valida+on

23 5. Road network comple+on Denver Depth image Intensity image Visual Quality: there are and roads and roads are extracted.

24 5. Road network comple+on Denver Depth image Intensity image Visual Quality: there are 10 ver+cal and 5 horizontal roads. 8 ver+cal and 4 horizontal roads are extracted.

25 5. Road network comple+on Oakland Depth image Intensity image Visual Quality: there are and roads and roads are extracted.

26 5. Road network comple+on Atlanta Depth image Intensity image Visual Quality: there are 7 ver+cal and 6 horizontal roads. 5 ver+cal and 4 horizontal roads are extracted.

27 Ques+ons