Research on pedestrian small target detection in dense scenes based on YOLOv8
Wenkao Yang, Dong Xueying · 2025
Pedestrian detection in dense scenes faces challenges like small target sizes and ambiguous scale features, which can result in missed detections and false positives. This paper presents an enhanced small target detection algorithm based on YOLOv8. The algorithm replaces traditional convolution with linear deformable convolution in the Bottleneck and integrates an ECA attention mechanism in the Neck, enhancing the detection efficiency for small targets. Experimental results confirm the effectiveness of the improved YOLOv8 algorithm. The proposed model performs exceptionally well on the dataset used in this study, achieving an AP50:95 score of 53.7%, indicating better accuracy at higher IoU thresholds. These results highlight the enhanced model's ability to detect small targets and its overall network robustness.