Deep Learning Based 3D Target Detection

Aiming at the problem of insufficient feature extraction capability in the current three-dimensional target detection algorithm based on LiDAR point cloud, this paper improves the three-dimensional target detection algorithm based on Pointpillars LiDAR. Firstly, the hybrid attention module is introduced to weight and perform feature fusion on pseudo-images, expanding the image sensing field through the weighting mechanism and enhancing the feature extraction capability of the algorithm in the 2D image module. Second, the point cloud processing module of Pointpillars is improved based on Swin Transform to better capture the features of the point cloud, process the relationship between each point and other point clouds in parallel, and better utilize the point cloud information. Finally, the algorithm is validated on the publicly available KITTI dataset with car class as the detection target, and it improves in easy, medium, and difficult cases, respectively, compared with the benchmark network. The experimental results show that the algorithm in this paper effectively improves the detection accuracy.