Deep Learning-Based LiDAR Perception for Traffic Cone Detection in Autonomous Formula Student Electric Vehicles

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate and efficient environmental perception is essential for autonomous vehicles. For autonomous Formula Student electric vehicles, reliable 3D traffic-cone detection is critical for track-boundary reconstruction and downstream path planning. However, the small physical size of traffic cones, sparse long-range LiDAR returns, and vegetation interference make this task challenging. This study proposes an efficient deep learning-based LiDAR detection method based on PointPillars. Specifically, three task-oriented components are incorporated into different stages of the PointPillars framework: a masked attention-pooling module is introduced into the Pillar Feature Network to improve sparse-point feature aggregation, Coordinate Attention is embedded into the bird’s-eye-view (BEV) backbone to enhance spatial feature representation, and a composite Smooth L1–3D Distance-IoU (DIoU) loss is designed to improve 3D bounding-box localization. Experiments were conducted on an in-house dataset collected using a 32-beam mechanical LiDAR mounted on a self-developed autonomous Formula Student electric race car. The dataset contains 1800 point-cloud frames and 9369 annotated traffic-cone instances. The proposed method achieved a BEV mean average precision (mAPBEV) and 3D mean average precision (mAP3D) of 86.40% and 77.24%, respectively, improving the original PointPillars by 3.08 and 4.52 percentage points, respectively. In the 15–20 m range, mAP3D increased from 64.67% to 72.46%. In additional on-vehicle experiments, the proposed method achieved approximately 29.8 FPS on the onboard industrial computer, while precision increased from 91.30% to 97.70% and the missed-detection rate decreased from 1.75% to 0.58% compared with the original PointPillars. These results indicate that the proposed method provides a favorable balance between detection accuracy and computational efficiency for LiDAR-based perception in autonomous Formula Student electric vehicles.