@article{HOU2025, 
author = {Yudi HOU and Hongchen YANG and Nengbin CAI},
title = {Research on a Fighting Behavior Detection Model in Videos Based on Improved YOLOv5s},
year = {2025},
journal = {Forensic Science and Technology},
volume = {50},
number = {3},
pages = {259-265},
keywords = {video reconnaissance, lightweight model, improved YOLOv5s, fighting behavior detection},
url = {https://www.sciopen.com/article/10.16467/j.1008-3650.2024.0043},
doi = {10.16467/j.1008-3650.2024.0043},
abstract = {With the wide application of surveillance systems, there is an increasing concern about public safety and security issues. Among them, the rapid detection and recognition of fighting behavior is very important for maintaining social order and security. However, traditional monitoring systems often face many challenges when dealing with large-scale video streams, including high computational complexity and resource-limited environments. In order to cope with these challenges, this paper proposes an improved fighting behavior detection model based on YOLOv5s, which reduces the number of parameters of the model and the computational complexity, so that the model can operate more efficiently in the resource-limited environment and detect various fighting behaviors more accurately. First of all, the open source interactive markup tool Labelimg was used to annotate the data set and train the network model with a large amount of data. Secondly, considering the need for rapid and accurate solutions in public security practice, lightweight network MobileNetv3 is used as the backbone network by comparing various convolutional structures to replace the original backbone network of YOLOv5s model, so as to reduce the number of parameters and calculation amount of the model and improve the model detection accuracy. By setting ablation experiments, the improved model is compared with other models and the original model. The experimental results show that compared with the original network, the detection accuracy of the improved model is increased from 92% to 94.4%, the computational load is reduced from the original 15.8 G to 3.1 G, and the detection speed of the algorithm can reach 0.153 s at the fastest, meeting the real-time requirements. And the detection accuracy is the highest among the three models. This model is suitable for public security practical application scenarios with high precision and limited memory and computing power.}
}