A Voxel-Based Skewness and Kurtosis Balancing Algorithm for Updating Road Networks from Airborne Lidar Data

Transcription

1 A Voxel-Based Skewness and Kurtosis Balancing Algorithm for Updating Road Networks from Airborne Lidar Data Li Liu & Samsung Lim School of Civil & Environmental Engineering University of New South Wales

2 Introduction Keeping road network database up-to-date is essential for many applications associated with our daily life. How to improve the accuracy of the geospatial database with limited resources. Updating road networks sometimes lags behind up to a few decades (Bentabet et al., 2003)

3 Introduction Current Road Network Updating Methods Field surveying : time-consuming, labor-intensive; Vector map comparison : generalization errors; Image processing : sensitive to the image resolution, not feasible in blocked areas; Lidar-based mapping : accurate, suitable for large-scale scenarios, cost-effective.

4 Skewness and Kurtosis Balancing Widely used in DEM generation, segmentation and classification; The assumption is that the measurements of a homogeneous class are supposed to assume a normal distribution (Duda et al., 2001), and the presence of other classes would distort the normal distribution.

5 Skewness and Kurtosis Balancing Test the effectiveness of intensity and elevation in skewness and kurtosis balancing; Figure 1. Results of elevation-based SKBA Figure 2. Results of intensity-based SKBA Figure 3. Results of SKBA based on the combination of elevation and intensity

6 Framework of The Algorithm Figure 4. Framework of Updating Road Network

7 Tiling of lidar points Skewness and kurtosis balancing is largely reliant on the assumption of normal distribution; The presence of multi-classes in lidar data and different road materials in a road network will distort the normal distribution; Vector data is utilised to buffer lidar points, segregating road points made of a similar material.

8 Partitioning of lidar points Lidar data typically consists of a huge amount of points. It is time-consuming to process the entire points without proper data management. Partitioning lidar points along the main direction of road segments according to the geographic coordinates.

9 Voxel-based filtering The intensity values of some trees in lowelevation areas also satisfy the normal distribution. It is essential to remove object points in lowelevation areas before the SKBA is applied.

10 Voxel-based filtering 1. Set the voxel size. 2. Calculate the voxel index for each point according to the coordinates and allocate points to the corresponding voxels. 3. Set the search index to the main direction of the road. 4. Check the bins along the search index for each tile. For each bin, find the index of a voxel in z direction that contains the point of the minimum z value in the height bin and remove the points whose voxel index in z direction is larger than the minimum. In addition, if the minimum voxel index in z direction is greatly larger than that of the neighbourhood, all points in the height bin are removed.

11 Skewness and Kurtosis Balancing 1. Calculate the skewness and kurtosis values of the dataset. 2. For a normal distribution, the kurtosis value is 3 and the skewness value is 0. If the absolute skewness is larger than the threshold and the absolute difference between kurtosis and 3 is larger than the threshold, go to step (3); otherwise, go to step (6). 3. The point with the maximum intensity value is regarded as the object point and is removed. 4. Recalculate the skewness and kurtosis values of the dataset. 5. Repeat steps (2)-(4) until all the points are processed. 6. If there are some remaining points, label them as candidates for road points.

12 Refinement of Candidate Road points (a) Road Candidates (b) Deletion of False Positives (c) Interpolation (d) Curve Fitting Figure 5. Procedures of Further Refinement

13 Experiments on SKBA Figure 6. Four airborne lidar datasets: (A) Anzac Parade, (B) Botany Street, (C) a residential area, (D) Barker Street near UNSW

14 Experiment Results (g) (a) (b) (c) (d) (e) Figure 7. Results of the extracted road network and respective centrelines: (a) extraction results of Anzac Parade, (b) estimated centrelines of Anzac Parade, (c) extraction results of Botany Street, (d) estimated centrelines of Botany Street, (e) extraction results of Barker Street, (f) estimated centrelines of Barker Street, (g) extraction results of a residential area, (h) estimated centrelines of the residential area. (f) (h)

15 Results Analysis Quantitative Analysis Table 1. Statistics of the Quantitative Analysis Sample 1: Residential area Sample 2: Anzac Parade Sample 3: Barker Street Sample 4: Botany Street Length of reference roads (m) Length of extracted roads (m) Length of undetected roads (m) Total length of Matched roads (m)

16 Results Analysis Quantitative Analysis

17 Results Analysis Quantitative Analysis

18 Results Analysis Quantitative Analysis Table 2. Results of the Quantitative Analysis Sample 1: Residential area Sample 2: Anzac Parade Sample 3: Barker Street Sample 4: Botany Street Completeness (%) Correctness (%) Quality (%) Weighted Positional Accuracy of Dataset (m)

19 Results Analysis Analysis of different voxel cell size (a) Voxel size of 2 m (b) Voxel size of 8 m (c) Voxel size of 10 m (d) Voxel size of 20 m Figure 8. Analysis of different voxel cell size

20 Concluding Remarks It is a practical method to update road network; Vulnerable to the problems of arcs of traffic islands; Rely on existing road networks to ensure the assumption of SKBA can hold. Omission error of road extraction may increase if multiple road materials appear in one tile.

21