Time-Sensitive 3D Single Target Tracking Based on Deep Hough Optimized Voting

Abstract

Aiming at the problem of time-sensitive single target tracking in 3D point cloud, a deep learning algorithm based on deep Hough optimized voting is proposed. Firstly, the algorithm uses PointNet++ network to calculate seed points and extract geometric features from template point cloud and search point cloud. A target-oriented feature extraction method is then used to encode the target information from the template into the search area. Secondly, potential target centers with high confidence are calculated and screened by seed point voting. Finally, multiple proposals are generated through sampling and aggregation of the target center points, and the proposal with the highest score is selected to generate a 3D target box. The algorithm can effectively avoid the time-consuming 3D global search, and is robust to the disorder, irregularity and sparsity of point cloud. In order to verify the effectiveness of the network, experiments are conducted on the public KITTI dataset. Experimental results show that the accuracy of the proposed network is improved by around 10%, compared to the current method based on 3D point clouds. At the same time, the method can run at 43.5 FPS on a single NVIDIA2080S graphics processor.

Keywords

• |time-sensitive target|single target|target tracking|point cloud|hough voting|deep learning|artificial intelligence