Research on real-time video target anomaly detection based on YOLO algorithm
Yifan Bian · 2025
In order to improve the accuracy and efficiency of real-time video target anomaly detection, the performance of the model based on YOLO algorithm in complex scenes is analyzed. By improving the feature extraction and multi-scale fusion strategy of YOLO, combined with spatio-temporal features and anomaly detection scores, the ability to recognize anomalous behaviors in dynamic scenes is significantly improved. The experimental results show that the model accuracy reaches 92.5% in light changing scenes, 88.7% in dense crowd scenes, and 85.3% in dynamic occlusion scenes. In terms of processing efficiency, the average processing time is 12.3 milliseconds and the frame rate is 81.3 fps for the light change scene, and the average processing time is 18.4 milliseconds and the frame rate is 54.3 fps for the dynamic occlusion scene. The performance of the model in complex environments verifies its efficiency and robustness.