MAT-PointPillars: Enhanced PointPillars algorithm based on multi-scale attention mechanisms and transformer

Abstract

Aiming at the problem that small and irregular detection targets such as cyclists have low detection accuracy and inaccurate recognition by existing 3D target detection algorithms, MAT-PointPillars (Multi-scale Attention and Transformer PointPillars), a 3D object detection algorithm, extends PointPillars with multi-scale vision Transformers and attention mechanisms. First, the algorithm employs pillar coding for semantic point cloud encoding and introduces an attention mechanism to refine the backbone’s upsampling process. Furthermore, the Transformer Encoder is introduced to improve the upsampling structure of the third stage of the backbone. On the KITTI dataset, our algorithm achieved 3D average detection accuracy (AP3D) of 81.15%, 62.02%, and 58.68% across three difficulty levels. Compared with the baseline model, the proposed algorithm improves AP3D by 2.44%, 1.19%, and 1.23% respectively. The real-time 3D object detection system is built based on ROS, and average running frames per second of the system is 22.63, which is higher than the sampling frequency of conventional LiDAR. By ensuring sufficient detection speed, the MAT-PointPillars algorithm can increase detection accuracy of cyclists in real-world scenarios.