18 October 2023 in IEEE Robotics and Automation Letters
Identifying moving objects is an essential capability for autonomous systems, as it provides critical information for pose estimation, navigation, collision avoidance, and static map construction. In this letter, we present MotionBEV, a fast and accurate framework for LiDAR moving object segmentation, which segments moving objects with appearance and motion features in the bird's eye view (BEV) domain. Our approach converts 3D LiDAR scans into a 2D polar BEV representation to improve computational efficiency. Specifically, we learn appearance features with a simplified PointNet and compute motion features through the height differences of consecutive frames of point clouds projected onto vertical columns in the polar BEV coordinate system. We employ a dual-branch network bridged by the Appearance-Motion Co-attention Module (AMCM) to adaptively fuse the spatio-temporal information from appearance and motion features. Our approach achieves state-of-the-art performance on the SemanticKITTI-MOS benchmark. Furthermore, to demonstrate the practical effectiveness of our method, we provide a LiDAR-MOS dataset recorded by a solid-state LiDAR, which features non-repetitive scanning patterns and a small field of view.
4 September 2024 in IEEE Transactions on Industrial Electronics
As a crucial issue for outdoor mobile robot navigation, place recognition (PR) remains challenging in long-term cross-scene applications, necessitating enhanced robustness of PR algorithms. In this article, a novel multimodal PR approach is proposed, which deeply fuses the camera and LiDAR to effectively compensate for their respective limitations. The introduced place descriptor consists of three branches: image-based, point cloud-based, and fusion-based. Specifically, the fusion-based branch employs a dual-stage pipeline, leveraging the latent correlation between the two modalities for information interaction and conducting channel-level fusion. Moreover, implicit alignment in the fusion branch, coupled with information supplementation from the two single-modal branches, better addresses the information loss caused by limited overlap in their field-of-view (FoV). In extensive experiments on Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI), North Campus Long-Term (NCLT), USVInland datasets, and the real-world campus test, the proposed method stands out in precision-recall curves with an average maximum F1 score of 0.949, demonstrating superior robustness and generalization compared with other state-of-the-art methods.