Study On Mobile Objects Detection And Tracking And Semantic Labeling Method Of 3D Point Cloud Scene

3D Lidar-based environment perceptional technology is the most significant approach in the domain of Unmanned Ground Vehicle(UGV).This approach holds the core status among almost every state-of-the-art UGV platform.This also indicates the future developing direction of Unmanned driving technology.Theoretically and practically,we solve two key problems from two aspects on some certain extent: first,mobile object detection and tracking methods in 3D point cloud scene;second,semantic labeling method for 3D point cloud scene.This dissertation studies on technology of environmental perceptional methods,and the main questions and work includes here:1.Given the practical application in on-board platform,we propose a simple and consecutive detecting scheme on 2D space grid map to extract the mobile objects in 3D point cloud scene.For a certain range of point cloud,it can extract the moving target.2.The moving target is tracked using an improved dynamic annealing(AD)method.Based on AD dynamic target tracking algorithm can perform very fast to search for the optimal posterior probability distribution of the state space,and give a more accuracy solution.For single target/single frame tracking task,this algorithm can be achieved in the millisecond level,while effectively reducing the tracking error at the same time.Experiments prove the method's validity and real-time performance.3.Proposed a new description structure for local 3D point cloud-Voxels neighbor structure(VNS).Based on the extraction of effective feature from point cloud,we construct VNS feature descriptor proposed in this paper.The preliminary semantic classification results are obtained by using supervising learning method-Random Forests.VNS feature descriptor combining with random forests method can preliminary obtain scene semantic labeling results more effectively.4.An point-wise spatial relationship is implemented based on actual contextual constraints of point cloud set.And further the conditional random field(CRF)model is constructed using this additional constraints.Finally,this dissertation uses graph cut to optimize the model and obtain further solutions.The method is proved to be able to further improve the basic energy network optimization results despite of some difficulty of classification condition.In order to prove our methods' efficiency and effectiveness,we arranged several experiments on dataset and offline real-time data.Based on the experimental results and analysis,our proposed tracking methods could achieve higher precision and do in realtime.Our method of semantic labeling in this dissertation could be compared to the stateof-the-art in the aspect of robustness.