Three-Dimensional Object Detection Network Based on Coordinate Attention and Overlapping Region Penalty Mechanisms

Wenxin Li, Shiyu Zhu, Hongzhi Liu, Pinzheng Zhang, Xiaoqin Zhang

Abstract: Three-dimensional target detection is a key technology in the fields of autonomous driving and robot control for applications such as self-driving cars and unmanned aircraft systems. In order to achieve high detection accuracy, this paper proposes a 3D target detection network with a coordinate attention training mechanism that generates voting feature points for better detection ability and an overlap region penalty mechanism that reduces false detection. In comparative experiments on public large-scale 3D datasets including the Scannet dataset and SUN-RGB-D dataset, the proposed method obtained an average detection accuracy mAP of 60.1% and 58.0% with an intersection ratio of 0.25, which demonstrates its superior effectiveness over the current main algorithms such as F-PointNet, VoxelNet and MV3D. The improved method is expected to achieve higher accuracy for 3D object detection relying only on point cloud information.