RAE3D: Multiscale Aggregation-Enhanced 3D Object Detection for Rail Transit Obstacle Perception
Lirong Lian, Zhiwei Cao, Yong Qin, Yang Gao, Wei Li, Jie Bai, Xuanyu Ge, Tangwen Yang
Abstract
Point cloud-based 3D object detection technology provides precise information for detecting obstacles in front of trains. In rail transit scenarios, obstacles are usually located at a distance, and the sparsity of point cloud data leads to information loss during feature extraction and spatial transformation, adversely affecting the accuracy of obstacle detection. To tackle this issue, we propose RAE3D, a single-stage end-to-end architecture that incorporates a multiscale voxel feature aggregation module on point queries (MVA-PQ) and a bird's-eye view multilevel auxiliary (BEV-MLA) module for efficient 3D object detection. The MVA-PQ module encodes multiscale voxel features into centroids of the final voxel layer, integrating spatial and scale information. The BEV-MLA module converts shallow 3D features into sparse 2D features using height compression, guiding the network to learn the spatial structure of small objects. We also introduce a 3D centroid offset loss for optimizing bounding boxes. Extensive experiments on the Rail3D and KITTI datasets have demonstrated the superiority of the proposed RAE3D.