A 3D object detection method for roadside LiDAR data based on an improved CenterPoint

Engineering notes

Although roadside LiDAR can provide a wide perception perspective to enhance the environmental perception capability of autonomous driving systems, issues such as sparse point clouds, target occlusion, and difficulties in detecting small targets significantly restrict the accuracy of traditional 3D object detection algorithms. Experiments on the DAIR-V2X-I roadside dataset demonstrate that the improved method achieves average precisions (AP) of 71.91%, 72.23%, and 68.61% for the Car, Pedestrian, and Cyclist classes, respectively.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Although roadside LiDAR can provide a wide perception perspective to enhance the environmental perception capability of autonomous driving systems, issues such as sparse point clouds, target occlusion, and difficulties in detecting small targets significantly restrict the accuracy of traditional 3D object detection algorithms. To address this, this paper proposes an improved CenterPoint method tailored for roadside LiDAR scenarios: First, a shape-aware data augmentation strategy is introduced, to enhance the model's robustness in recognizing incomplete objects by simulating various occluded states through random removal of local points, cross-object point swapping, and farthest point sampling. Second, a sparse residual 3D backbone network is designed by integrating residual connections into sparse convolutional modules, strengthening the retention and propagation of sparse point cloud features and mitigating information loss in deep networks. Third, a Spatial Semantic Feature Aggregation (SSFA) module incorporating Self-Calibrated Convolutions (SCConv) is proposed to adaptively fuses high-level semantic features and low-level spatial features from the BEV perspective, thereby enhancing the perception of both local details and global structures. Experiments on the DAIR-V2X-I roadside dataset demonstrate that the improved method achieves average precisions (AP) of 71.91%, 72.23%, and 68.61% for the Car, Pedestrian, and Cyclist classes, respectively. Compared with the original CenterPoint, these results represent improvements of 2.56%, 2.84%, and 4.06%, verifying the effectiveness of the proposed improvement strategies in complex roadside environments.

Links and sources

• Official / arXiv page
• PDF from original source

Comments