YOLO_NMV: intelligent pedestrian detection for non-motorized vehicles

Non-motorized vehicles are widely used for short-distance travel and logistics, but distractions or illegal behavior often lead to accidents, posing a threat to pedestrians and drivers. To enhance the intelligence and safety of non-motorized vehicles, efficient pedestrian detection algorithms are needed. Based on the YOLOv8 architecture, the YOLO_NMV algorithm is proposed. The improvements include introducing StarNet_ADown in the backbone to enhance feature extraction. In the Neck layer, the C2f operation is improved using StarBlock, while LDTConv (Linear Deformable T-Convolutions) is employed in a different module to reduce channel redundancy and improve model robustness. Compared to the YOLOv8n model, YOLO_NMV reduces parameters, computation, and weight by 40.9%, 31.7%, and 38.1%, respectively, while improving precision, recall, mAP50, and mAP50-95 by 5.4%, 1.4%, 2.4%, and 2.3%. YOLO_NMV is more lightweight, accurate, and robust, making it suitable for deployment in non-motorized vehicles with limited computing power.