An Improved RDE-YOLOV10 Algorithm for Campus Conflict Behavior Recognition

In today’s society, technologies for detecting fights and violent altercations are crucial for preventing conflicts and ensuring public safety. Therefore, this paper proposes an improved fight detection algorithm based on YOLOv10. First, the model's backbone is replaced with RepViT, a design that facilitates the capture of multi-scale contextual information. The C2f_DBB focus module is then incorporated, in which the Bottleneck’s convolutional layers of the original C2f module are replaced with Diverse Branch Block (DBB), leading to enhanced feature extraction capability. Finally, the EIoU loss function is adopted to further enhance localization accuracy. Experimental results demonstrate that the proposed algorithm achieves a 3.6% improvement in and a 1.2% improvement in :0.95 compared to YOLOv10n, effectively meeting the real-time detection requirements of fight and altercation behaviors in critical areas.