An Improved YOLO V5-s Algorithm for Real-Time Pedestrian Detection in Crowded Public Scenes
Abstract
Pedestrian detection is critical in fields such as autonomous driving, surveillance, and public safety, especially in crowded environments where occlusion and overlap of individuals pose significant challenges. While deep learning-based approaches have achieved notable progress, real-time detection on mobile devices remains difficult due to computational constraints. This paper proposes an improved YOLO V5-s algorithm designed for efficient pedestrian detection in crowded scenes on mobile platforms. To enhance performance, ShuffleNetV2 is introduced as a lightweight backbone, reducing model complexity while maintaining detection accuracy. The Convolutional Block Attention Module (CBAM) is integrated, employing channel and spatial attention mechanisms to optimize feature extraction. Additionally, data augmentation techniques and label smoothing are used to expand the training dataset, and the Complete Intersection over Union (CIoU) loss function is incorporated to improve detection precision and reduce missed detections. Experimental results on the MOT20det dataset show that the enhanced algorithm achieves a mean Average Precision (mAP) of 0.983 and a frame rate of 144 FPS, outperforming the original YOLO V5-s in both speed and accuracy. The lightweight model is suitable for real-time pedestrian detection on mobile devices in densely populated areas.