Research On 3D Object Tracking Algorithm For Autonomous Driving

3D object tracking is a long-term research task in the field of autonomous driving.Since lidar can generate more accurate point cloud data,it has become one of the most commonly used sensors in the field of autonomous driving.Different from image data,due to the disorder,sparsity and unstructured characteristics of point cloud data,it brings greater challenges to the design of 3D object tracking algorithms.In this paper,3D object tracking for autonomous driving is explored and researched,which can be categories single object tracking and multi-object tracking.1.Frustum-based Double Siamese Network for 3D Single Object Tracking.Aiming at the sparse point cloud data and wide distribution range,the proposed frustum-based double Siamese network for 3D single object tracking can combine the advantages of image and point cloud data to make up for the defects of point cloud data.The proposed frustum-based region proposal module can leverage the image tracking results to obtain a more accurate 3D search space.In addition,a rapid accuracy verification strategy is designed by using the property of continuity of 3D tracking to verify the accuracy of 2D tracking results,which solves the defect of the influence of the previous-stage results on the subsequent-stage results in the serial network structure.Experiments in the KITTI tracking data set show that the proposed method has certain advantages compared with the state-of-the-art(SOTA)tracking methods in the same period.2.Target-aware Voxel-based 3D Siamese Tracker for Point Clouds.Aiming at the existing single object tracking methods based on point cloud are difficult to effectively generate can-didates and combine mature visual tracking experience,we propose a novel target-aware voxel based 3D Siamese tracker.Using the characteristics of structure of voxel grid,we can easily learn from mature visual tracking experience and propose the Siamese region proposal network for quickly generating candidates.Then,with the help of the proposed multi-scale Region of Interest pooling operation,multi-scale semantic features are extracted to further refine the tracking results.In addition,the attention module proposed in this paper aggregate template features and search region features,which effectively improves the representation of features.Experiments in the KITTI data set show that the proposed method has certain advantages compared with the SOTA tracking methods in the same period.3.Siamese network based 3D multi-object tracking.In order to solve the problems that the motion prediction module cannot use the semantic features of the data and is too dependent on the detection results in 3D multi-object tracking framework,the single-object tracking proposed in the previous article is extended to the multi-object tracking framework,and a 3D multi-object tracking algorithm based on the Siamese network is proposed.Based on the 3D Siamese tracker,two motion prediction modules are designed to explore the impact of motion prediction models on the performance of multi-object tracking algorithms,and a similarity measurement strategy combining angle information and space information is designed.The results in the public test sets of nu Scenes and KITTI show that the proposed method surpasses the point-based SOTA 3D multi-object tracking methods in the main met-ric.