Curb Point Clouds Extraction from Vehicle-Borne Laser Scanning Data
Haifeng LUO,Lina FANG,Chongcheng CHEN
DOI: https://doi.org/10.3724/SP.J.1047.2017.00861
2017-01-01
Abstract:Vehicle-borne laser scanning system can quickly and accurately obtain 3D point cloud data of the street scene by multiple sensors and data processing technology. Extracting street curb point clouds from vehicle-borne laser scanning data is important for many applications such as road panning and maintenance, but it is difficult to directly extract curb point clouds from the original data due to the high point density, large amount of data, uneven spatial distribution of point clouds, object obscuring each other and complex urban street scenes. A novel algorithm for extracting curb point clouds from vehicle-borne laser scanning data based on Support Vector Machine(SVM)is proposed in this paper. The algorithm is described as followings:firstly, the original data is thinned and segmented into a series of point cloud blocks to improve efficiency. Secondly, a point clouds feature vector is constructed including the relative elevation, normal vector direction, multi-scale height difference and height variance, through the analysis of spatial distribution and local geometric features of curb point clouds. Thirdly, a new method is proposed to refine the point cloud feature of the border of ground and object on the ground such as point clouds of the bottom of tree, building facade, fence, and so on, which avoids errors in dividing the border point clouds into curb point clouds. Fourthly, the obtained training samples are processed by artificial choice. Radial Basis Function(RBF)is selected as kernel. Particle Swarm Optimization(PSO)is used to optimize penalty factor C and kernel function parameter γ, then the feature vector is taken as the input to train SVM. The obtained SVM is used to extract curb point clouds from original data. Finally, cluster analysis is performed on the extracted results for eliminating noisy point clouds when the number of point clouds in cluster is less than threshold. Experiments were undertaken to evaluate the validities of the proposed algorithm with four different street scene datasets acquired by Street Mapper 360 System and Lynx Mobile Mapper System, respectively. The completeness of the results of curb point clouds extraction is over 94.99%, correctness is above 91.88%, and quality is above 90.55%. This proves that the proposed algorithm not only has high accuracy, but also has strong robustness to extract curb point clouds with regular or irregular shape from complex urban street scenes of vehicle-borne laser scanning data.