Semi-Supervised Roadside Vehicle Point Cloud Detection and Few-Shot Fine-Tuning Method Based on SAM

3D object detection using roadside LiDAR is a key technology for achieving large-scale environmental perception in autonomous driving. It plays an important role in various applications such as road monitoring, vehicle-to-everything (V2X) systems, and perception assistance. However, significant variations in LiDAR models, installation methods, and road environments introduce considerable domain gaps, making conventional detection networks highly dependent on large amounts of annotated data to achieve satisfactory performance. This paper explores the potential of the Segment Anything Model (SAM) in point cloud detection tasks and proposes a novel approach: a bird’s-eye view (BEV) feature map is first generated by integrating the spatial distribution and reflectance intensity of the roadside point cloud; then, high-frequency texture and edge features are incorporated into SAM, and few-shot fine-tuning is performed to obtain segmentation masks on the BEV feature map; finally, the segmentation results are combined with the original point cloud to generate 3D detection boxes. The proposed method is evaluated on the DAIR-V2X-I dataset, and the experimental results demonstrate that its performance is close to traditional networks trained on large datasets, showing strong generalizability and promising real-world applicability.