An Efficient Real-Time Object Detection and Tracking Framework Based on Lidar and Ins

Obstacle detection and tracking is an integral part of the autonomous vehicle perception algorithm. Because most of the existing tracking algorithms have the problems of insufficient accuracy and poor real-time performance, an obstacle detection and tracking method based on LiDAR is proposed in this paper. Firstly, for the scene with the undulating ground in the environment, a twice-ground segmentation method based on plane fitting and scan line geometric features is proposed to accurately and robustly extract the high obstacle point cloud. Secondly, the density clustering algorithm is optimized, and a convex hull rectangular 3D bounding box fitting algorithm is proposed to detect obstacles. Finally, the Mahalanobis distance measurement feature is used to realize the data association between the previous and the current frame. And the interacting multiple model filter algorithm embedded in the unscented Kalman filter is used to estimate the state of the object optimally. Based on the public data set, the proposed algorithm improves the accuracy of tracking detection. After verification on the self-developed real vehicle experimental platform, the results show that the algorithm has good object tracking and correlation performance.