LiDAR Curb Detection Method Based on Distance-Adaptive Feature Enhancement and Dynamically Balanced Loss

In terms of environmental perception in autonomous driving, high-precision curb detection is actually a particularly crucial task to ensure that vehicles can plan their paths and drive safely. In complex urban road scenarios its criticality becomes even more evident. Currently, the existing curb detection methods based on LIDAR have encountered a rather prominent challenge in long-distance curb detection because long-distance curbs suffer from insufficient feature expression due to sparse point clouds, resulting in frequent missed detections. To solve this problem, this paper proposes DA-CurbNet, a curb detection framework integrating distance-adaptive feature enhancement and dynamically balanced loss. Specifically, this module can enhance the distance features by dynamically matching the scale of the convolution kernel with the distance interval. In addition, in order to increase the transmission weight of long-distance features, this paper also introduces a distance-aware skipping mechanism and adopts a dynamic balance loss function to make the model pay more attention to the remote suppression performance. Experiments conducted on the 3D-Curb dataset show that DA-CurbNet significantly improves the detection effect of long-distance curbs. Compared with CurbNet (SOTA deep learning methods), the recall rate of DA-CurbNet has increased across all distance intervals, the D3 interval achieves the largest recall increase (5.34%), and the D4 interval gains the maximum F1-score improvement 0.0292. Globally, DA-CurbNet maintains a competitive total precision (87.32%), with total recall rising by 2.36% and total F1-score optimized to 0.8807. This work provides a relatively reliable solution for long-distance curb perception in autonomous driving.