Based on deep learning: campus security and abnormal behavior early warning

Abstract

To address large target scale variations, severe occlusions in crowded scenes, and complex backgrounds in campus surveillance, this study proposes an abnormal behavior detection model named B-YOLOv10-WH. The model introduces the Shape-IoU loss function to improve bounding-box regression accuracy, the Wavelet Transform Convolution (WTConv) module to enhance multi-scale feature extraction, and a Histogram Transformer module to strengthen global scene understanding and local feature capture. Experimental results show that B-YOLOv10-WH achieves Precision, Recall, F1-score, and mAP values of 92.07%, 92.01%, 92.04%, and 96.68%, respectively, improving by 4.67%, 5.91%, 5.30%, and 4.02% over the original YOLOv10. In addition, the number of network layers is reduced by 15.67%.The proposed model improves detection accuracy, robustness, and real-time performance in campus abnormal behavior detection tasks. It also demonstrates strong practical value and deployment potential, providing effective technical support for intelligent campus security monitoring and early-warning systems.