A PointPillars-based 3D point cloud object detector of USVs for small target detection in dynamic aquatic environments

Abstract

LiDAR, a crucial sensor for Unmanned Surface Vehicles (USVs), allows for precise 3D modelling but encounters challenges in real-time target detection due to sparse point clouds. Current 3D point cloud detectors struggle to effectively capture fine-grained details and dynamic water surface features, while high-performance models often rely on custom operators, making deployment more complicated. Additionally, current water surface datasets lack the resolution necessary for small target detection. To tackle these issues, this study enhances the PointPillars model with the Voxel-Guided Label Assignment (VGLA) strategy, improving feature extraction through adaptive label assignment. A high-resolution point cloud dataset focused on small aquatic objects has also been developed based on 128-beam LiDAR. The proposed PointPillars-VGLA achieves 3D AP scores of 89.50%, 83.70%, and 75.20%, as well as BEV AP scores of 95.20%, 91.00%, and 86.70% across three target categories. Ablation experiments confirm the effectiveness of the VGLA module, with accuracy gains of up to 2.27% over CenterPoint. Deployed on the Jetson AGX Orin with TensorRT, the model achieves real-time inference at 30 FPS, enabling efficient detection and tracking in dynamic aquatic environments.