Segmentation And Classification For 3D Point Cloud Data Of Outdoor Scene

With the rapid development of 3D rangefinder technology,3D point cloud data have been applied to many application domains, such as autonomous navigation, industrial inspection, and reverse engineering. In the processing techniques of 3D point cloud data, segmentation and classification are two key techniques, especially for 3D point cloud of outdoor scene, and they are the foundation of environmental analysis and autonomous navigation. In real outdoor scene, it is a challenging job to design the effective methods to achieve the segmentation and classification due to complex structures and strong noise. In order to improve the accuracy and robust of the segmentation, this paper will do an in-depth research on the geometrical features, topological structures, and neighborhood characteristics of outdoor scene, and then the reliable classification is achieved based on the segmentation.In the segmentation part, this paper proposes a new approach for segmenting the 3D point cloud data of the outdoor scene based on the feature ball, and it uses the idea of the gradual segmentation to complete the task. First, the geometrical features are used to divide the point cloud into three parts:scatter points, linear points and surface points, and a feature ball is constructed which consists of the scatter player, tangential layer, and normal layer. Then, the mean shift algorithm is used to cluster the normal layer points to segment the surface points into several surface regions. If the normal vector and elevation of some surface region meet the certain thresholds, it is chosen as the component of the primary ground. In order to get the whole ground, some points which are close to the primary ground will be merged into it. Finally, the non-ground points are segmented by the DBSCAN method. In the experiments, the performance of the proposed segmentation method is evaluated by using four groups of 3D point cloud data, and the results show that the method is accurate and robust.In the classification part, this paper presents a novel method for classifying the 3D point cloud data of the outdoor scene by using the conditional random field with high-order cliques. The conditional random field contains three types of models:the point model constituted by a single point, the edge model constituted by two adjacent points, and the high-order clique model constituted by some similar points, In order to construct the high-order clique models, the feature ball is used to segment the surface points and linear points, and then the K-means cluster algorithm is used for each segmented part and the scatter points, and the high-order clique models are obtained. Then the feature vectors of the point model, edge model, and clique model are computed, and a conditional random field model for the outdoor scene is constructed by using the feature vectors and labels of training samples. The optimal parameters of the conditional random field model are got by the subgradient algorithm. For the real 3D point cloud data of the outdoor scene, we use the graph-cut algorithm to obtain the classification results, In the experiments, two datasets are applied to evaluate the proposed classification method, and the results show that the method has the superior performance in many respects such as recall rate and accuracy.