GSW-Yolo: Improved Light-Weight Person Detection Method Based on YOLOv8

In this study, we propose GSW-YOLO, an enhanced lightweight person detection method based on YOLOv8, designed to improve detection accuracy and computational efficiency. The method incorporates Ghost Modules to reduce model parameters, Shuffle Attention Mechanism to enhance feature representation, and Wise Intersection over Union (WIOU) to improve bounding box evaluation. Experiments conducted on COCO and Pascal VOC 2012 datasets demonstrate that GSW-YOLO consistently outperforms YOLOv8 and its variants. On the COCO dataset, GSW-YOLO achieves an mAP50 of 0.6702 and mAP50-95 of 0.4308, while on Pascal VOC, it records an mAP50 of 0.7774 and mAP50-95 of 0.5323, both with low GFLOPs of 6.264 and fast inference speeds of 3.741 ms and 3.509 ms, respectively. Ablation studies reveal that the combination of Ghost Modules, Shuffle Attention Mechanism, and WIOU contributes significantly to these results, delivering higher precision and recall while maintaining lightweight efficiency. With these characteristics, GSW-YOLO is well-suited for real-time applications, including public surveillance and human activity tracking on resource-constrained devices. This research highlights the potential of GSW-YOLO to deliver reliable and scalable person detection solutions for real-world scenarios.