Frus-PointPillars-based method for distant target detection in roadway point clouds

In in-vehicle intelligent driving systems, although LiDAR–camera fusion can acquire 3D information, it remains insufficient for detecting small distant objects beyond 50 m due to sparse point clouds and limited feature representation. Existing detection methods mostly focus on close-range targets and do not adequately correspond to the challenges of small pixel occupancy, sparse point cloud, and easy confusion of small targets in the distant view. We propose a Frus-PointPillars-based algorithm that projects 2D detection boxes into 3D frustums to narrow the search scope and suppress background, employs dual pooling in the voxel feature extractor to retain both global and local information, and integrates a multi-scale residual graph-convolution fusion module with a global-aware attention mechanism. On KITTI, our model achieves 68.78%/64.17%/56.43% 3D mAP for Car/Cyclist/Pedestrian and 47.01% long-range (|x|≥40m) mAP (+3.81% versus PointPillars). On nuScenes, 3D mAP rises by 5.28% at 31FPS, demonstrating enhanced accuracy and robustness for distant small-object detection.