Research on 3D target detection algorithm for laser point cloud based on VP-SECOND network

Abstract Laser point cloud target detection algorithms are not sufficient for target feature extraction in complex scenes, which can result in low detection accuracy of long-distance targets and occluded targets. To address this problem, this paper proposes a three-dimensional (3D) target detection network VP-SECOND, which integrates Voxel self attention assisted network(VSAA) and Polarized self attention(PSA). The network takes SECOND as the basic network structure, firstly, in the voxel coding stage, a deformable voxel feature coding module is designed to replace the traditional coding module; Secondly, in the feature extraction stage, Transformer is applied to the voxel level, and a voxel Transformer network is designed to have a unique coding capability and to be able to expand the receptive field. Transformer network with a unique coding capability and the ability to expand the receptive field, which effectively improves the efficiency of searching relevant voxels by fast voxel querying; Finally, PSA is introduced in the 2D backbone part, which enables the model to learn more nonlinear detail semantic information and improves the detection accuracy of the network in small targets. The experimental results on the KITTI dataset show that compared with the benchmark algorithm, the mAP of VP-SECOND for cars, pedestrians, and cyclists is improved by 8.40%, 6.14%, and 13.12%, respectively, which proves the effectiveness of the VP-SECOND algorithm, and effectively improves the detection performance of the model at long distances and in complex scenes.