Research On Object Detection For High Resolution 3D Point Cloud

With the development of science technology in recent years, the manners of human perceiving world have been developed from 2D optical imaging technology to 3D laser scanning technology. After the development of optical, infrared, microwave, multispectral and hyperspectral systems, 3D laser scanning systems have become the most important remote sensing sensors. The high resolution 3D point cloud accquired through 3D laser scaning system has the properties of huge amount of data, irregular and cluttered scenes. The research for high resolution 3D point cloud has become a research topic in recent years. Object detection is the foundation of scene understanding, and it provides underlying object and basis of analysis for high level scene understanding. Research on object detection for high resolution 3D point cloud has the challenges both in theory and in application. This thesis investigates the problem of object detection for high resolution 3D point cloud of complex scenes, and analyzes the drawbacks of existing 3D point cloud object detection methods. The main contributions of this thesis are the proposal of Hough forest based 3D point cloud individual object detection model, and the proposal of saliency based 3D point cloud road boundaries detection model.(1) For 3D local patch extraction, based on the two properties of supervoxel over-segmentation algorithm: cannot acrossing object boudndaries and high computational efficiency, the thesis proposes a supervoxel neighborhood based 3D local patch extraction method. To ensure the best distinctiveness and computational efficiency of extracted 3D local patches, the thesis analyzes the distinctiveness under different sizes of supervoxel nieghborhood based on Laplace-Beltrami scale space(LBSS) theory. The proposed method covers the drawbacks of exisiting local patch extraction methods: unconsidered the object shape structure and cannot controlling the sizes of extracted local patches.(2) For class-specific individual object detection, the thesis proposes a 3D point cloud object detection model based on distance weighted circular voting Hough forest. The model extracts local patches from 3D point cloud based on octree, and describes loca pathces through geometric structural features and spectral features which ensure point-density invariance and rotation invariance. To detect rotated objects in complex scenes, a generalized Hough voting method based on distance weighted circular voting is proposed. The proposed method covers the limit of weak extension or the limit of dependence on the individual object shape completeness of existing methods. Experimental results demonstrates the effectiveness and robustness of proposed method on complex scenes such as occlusion, overlap and rotation.(3) The thesis proposes a 3D point cloud individual object detection method based on supervoxel neightborhood with Hough forest framework. The proposed method covers the limit of unconsidered object shape structure in scenes and the limit of at the cost of a low false positive rate of the previous distance weighted circular voting Hough forest method. By extracting 3D local patches based on supervoxel neighborhood, and achieving rotation invariance based on combination of local reference frame and circular voting, the proposed method achieves higher detection performance and computaiton efficiency.(4) For continuous 3D point cloud object detection, the thesis treats road boundaries in urban scenes as the research object, and proposes a local normal saliency based road boundaries detection method. First, the thesis proposes a normal based 3D point cloud saliency computational model. Based on the combination of saliency description and relative spatial distribution in road context, a road structure detection method is then proposed. Finally, a PCA based projection method is proposed to accurately extract road boundary points. Experimental results demonstrate that the proposed method achieves high detection performance and computational efficiency.