Research on Crowd Target Counting Method based on YOLOv5 Algorithm

With the intensification of social activities, public management and public safety have ushered in huge challenges. Aiming at the problem that manual guards can easily overlook abnormal events, this paper designs a crowd counting algorithm based on YOLOv5 (YOLOv5-AT), which introduces the attention mechanism into the detection network to enhance the feature learning of small targets in the image. The feature map reorganization method is used to enhance the richness of feature information, thereby improving the accuracy of detecting the population. The experimental results show that compared with YOLOv5, the MAE and RMSE of the personnel number detection task are reduced by 38.6 and 41.7. While ensuring the real-time detection, the algorithm effectively improves the detection accuracy, which has important practical significance for public safety incidents and commercial analysis.