DSR-YOLO: A lightweight and efficient YOLOv8 model for enhanced pedestrian detection

Abstract

This paper presents DSR-YOLO, a pedestrian detection network that addresses critical challenges, such as scale variations and complex backgrounds. Built on the lightweight YOLOv8n architecture, it incorporates DCNv4 modules to enhance the detection rates and reduce missed detections by effectively learning key pedestrian features. A new head component enables detection across various scales, whereas RFB modules improve accuracy for smaller or occluded objects. Additionally, we enhance the initial C2f layers with a modified block that integrates SimAM and DCNv4, minimizing the background noise and sharpening the focus on the relevant features. A second version of the C2f block using SimAM and standard convolutions ensures robust feature extraction in deeper layers with optimized computational efficiency. The WIoUv3 loss function was utilized to reduce the regression loss associated with bounding boxes, further boosting the performance. Evaluated on the CityPersons dataset, DSR-YOLO outperformed YOLOv8n with a 14.9 % increase in mAP@50 and 6.3 % increase in mAP@50:95, while maintaining competitive FLOPS, parameter counts, and inference speed.