EfficientYOLO-PV: A Lightweight and Efficient Object Detection Algorithm for Pedestrians and Vehicles

Addressing the issues of missed detections and false alarms in pedestrian and vehicle detection for intelligent transportation and autonomous driving, along with the demand for model lightweighting, this paper proposes EfficientYOLO-PV, a lightweight algorithm based on improved YOLOv8. Through four key innovations: The backbone network employs the novel PFM-SFENet, integrating enhanced FasterNet-Blocks with a dynamic ECA attention mechanism to strengthen multi-scale feature extraction while reducing computational load;The C2fPDA adaptive deformable convolution module is introduced to improve adaptability to target deformation and occlusion; A Branch-Reinforced Feature Fusion (BRFF) module is proposed, embedding lightweight LSKA to augment spatial perception and utilizing channel attention for deep semantic fusion; A geometrically and angularly dual-constrained loss function (GVALoss) is designed to optimize bounding box regression. Experiments demonstrate that compared to YOLOv8, the proposed algorithm achieves a 7.5% increase in mAP @ 0.5, a 43.5% reduction in computational complexity (FLOPs), and a 49.3% decrease in parameter count. Among eight comparative models, it exhibits optimal performance in both detection accuracy and speed, achieving comprehensive improvements in precision, efficiency, and lightweighting.