Design of pedestrian detection algorithm for new energy intelligent connected vehicles based on the fusion of YOLOv8 and millimeter-wave radar

Pedestrian detection for new energy intelligent connected vehicles faces dual challenges of accuracy and real‑time performance, while single sensors exhibit poor adaptability to complex scenarios including nighttime, rain, fog and occlusion. This paper proposes a pedestrian detection algorithm based on deep fusion of YOLOv8 and millimeter‑wave radar. First, for the visual branch, we reconstruct the C2f module with lightweight channel‑space joint attention and introduce a high‑resolution P2 detection head to strengthen small‑target feature representation. Second, a cross‑modal deformable attention (CMDA) module is developed to compensate geometric mismatches between sparse radar point clouds and visual features via learnable spatial offsets. Different from conventional deformable attention, CMDA adopts heterogeneous cross‑modal query‑value asymmetry, physics‑constrained offset generation derived from radar range‑velocity priors, and sigmoid‑gated adaptive modal weighting. SIoU loss and augmented complex‑scene samples are utilized in training to enhance model generalization. Evaluated on nuScenes, VoD and self‑built datasets, the proposed algorithm achieves 79.4% pedestrian AP and obtains 76.9% AP under nighttime conditions. With 14.3 M parameters, it runs at 38.6 FPS on NVIDIA Jetson AGX Orin and satisfies in‑vehicle real‑time requirements. This work provides an engineering‑feasible solution for pedestrian safety perception and offers methodological references for multimodal deep fusion paradigms.