Research on Detecting Violations through Fences based on Improved YOLOv11

Abstract

With the continuous advancement of intelligent security and intelligent video surveillance technologies, deep learning-based abnormal behavior detection has gradually emerged as a critical research direction in the security field. Traditional manual patrols and conventional video surveillance systems struggle to achieve rapid and accurate detection and early warning. To address the issues of missed detections, false detections, and low detection accuracy in pedestrian fence-climbing detection under complex scenarios such as occlusion, dense multi-target conditions, and simultaneous multi-person climbing, an improved YOLOv11-based illegal wall-climbing detection algorithm named ban-YOLOv11 is proposed. In the backbone module, GhostNet modules are combined with convolutional modules to effectively reduce computational cost and parameter count, enabling the model to be deployed on edge devices. Furthermore, the Convolutional Block Attention Module (CBAM), a hybrid-domain attention mechanism, is introduced at the end of the backbone network to mitigate interference from irrelevant targets in complex backgrounds, fuse multi-scale information, and achieve more comprehensive feature aggregation. Experimental results demonstrate that the proposed algorithm achieves a mean average precision (mAP) of 98.0%, which is 2.4% higher than the original model, and a detection speed of 80.6 frames per second (FPS) while reducing the parameter count. It effectively alleviates the problems of missed detections, false detections, and low detection accuracy, while maintaining high real-time performance, making it more suitable for real-time pedestrian fence-climbing detection.