An efficient image-guided-based 3D point cloud moving object segmentation with transformer-attention in autonomous driving

Abstract

For intelligent transportation systems, moving object segmentation (MOS) provides valuable information for robots and intelligent vehicles, such as collision avoidance, path planning, and static map construction. However, all existing 3D point cloud MOS methods are based on LiDAR-only, which limits the ability to fuse supplementary information from different sensors. In this article, we solve the robust and accurate 3D MOS problem by designing a dual-stream network that integrates point clouds and images. We propose a perspective residual mechanism to mine the spatio-temporal motion information of point clouds, and design a fusion module based on Transformer Attention to extract multi-scale feature information from point clouds and images, improving the segmentation integrity of moving objects. Many experiments on the benchmark dataset show the superiority of our method. On the Semantic-KITTI, we outperform the advanced method by 6.5% mIoU. And we further apply our proposed model to the Semantic-KITTI: Moving Object Segmentation competition and achieve an advanced ranking on the leaderboard.