Improved Pedestrian Detection Algorithm for Traffic Vehicles Based on YOLOv8

This paper proposes an enhanced YOLOv8 framework for traffic object detection, specifically optimized for small and occluded targets in complex urban environments. The core improvements include: (1) SPD-Conv modules replace strided convolutions to preserve spatial features of blurred or small objects; (2) BiFormer dynamic sparse attention for background suppression; (3) an auxiliary 160x160 detection head for small targets; (4) N-IoU loss optimizing localization for overlapping objects. Evaluations on BDD100K demonstrate significant accuracy improvements: [email protected] increases from 54.76% to 63.64% (+8.88pp) and [email protected]:0.95 from 32.62% to 40.16% (+7.54pp), while maintaining low complexity (3.8M parameters). Validation on KITTI confirms improved generalization, achieving 94.44% [email protected]. The balanced accuracy-efficiency trade-off makes this framework practical for autonomous driving deployments.