| The 3D point cloud data collected by 3D lidars have become a significant resource for autonomous vehicles to acquire road information.Among the environment perception algorithms based on point cloud data,obstacle detection and dynamic obstacle tracking algorithms play a vital role in the architecture of autonomous vehicles.However,due to the turning or moving of the autonomous vehicle when the lidar works,the collected data often cannot truly reflect the environmental information.And the traditional Euclidean clustering algorithm often causes false detection in the vicinity or missed detection in the distance if the Euclidean distance threshold is not selected properly.If a better solution can be found,the reliability and real-time performance of lidar detection and tracking will be greatly improved.This paper proposes a set of feasible obstacle detection and tracking algorithms to solve these problems from four main aspects:First,a de-distortion algorithm is applied to diminish the influence of distortion caused by the moving and turning of lidar.Second,we applied an adaptive threshold of Euclidean distance in the improved algorithm,so the improved algorithm is able to detect the relatively small objects in distance while it can also detect the objects nearby without misjudgment.Third,the algorithm also optimizes the real-time performance of the detection algorithm in areas such as neighborhood search algorithms and point cloud segmentation.Finally,we also enables the algorithm to better implement the tracking function by introducing new similarity calculation strategies.These features are all confirmed by the results of experiments which show that the algorithm can complete the detection and tracking of multiple obstacles with high accuracy and running speed.At the same time,with the help of proposed method the distortion of point cloud reduced,the detection distance has increased by nearly 8 meters and the clustering speed has also increased by 15%. |
PDF Full Text Request