The Optimization and Experimental Analysis of Improved YOLOv5 Algorithm for Small-Scale Pedestrian Detection

Abstract

In intelligent security monitoring, small-scale pedestrians (≤32×32 pixels) face challenges of low detection accuracy and high missed detection rates due to scarce features and severe background interference. This study proposes an enhanced YOLOv5 algorithm with three synergistic optimizations: (1) integrating a lightweight Channel-Spatial Attention Module (CBAM) into PANet to amplify small-target features; (2) optimizing anchor boxes via K-means++ clustering to improve scale adaptation by 12.3%; (3) combining transfer learning and Mosaic augmentation to strengthen generalization. Experimental results on COCO, Caltech, and a custom campus dataset show the improved model achieves 6.8% higher mAP@0.5 (45.5%) on COCO and 7.5% on Caltech, with FPS >45. Ablation studies confirm CBAM contributes the largest gain (2.5%). This research offers a robust solution for small-scale pedestrian detection in intelligent security systems.