Aiming at the problems of existing sensor-based intrusion detection techniques such as high false alarm rate and secu-rity risks,an improved area intrusion detection algorithm for YOLOv5 was proposed.Based on YOLOv5,the CBAM attention mechanism was introduced in Backbone to enhance the network's feature extraction ability;the operation layer was added in Neck to continue the up-sampling processing of feature maps,and the feature maps obtained after the operation were Concat fused with the feature maps of the second layer in Backbone,so as to obtain larger feature maps for small target detection;the mask method was combined with the image pixel coordinate system to delineate the alert area and detect suspicious targets enter-ing the alert area in order to prevent illegal intrusion.The experimental results show that the intrusion detection algorithm has a mAP value of 83.4％,which is 1.8％,17.3％,28.2％,and 40.6％higher than YOLOv5,YOLOX,SSD and Faster-RCNN respec-tively,and reaches a detection speed of 25.4 frames/s,which is second only to YOLOv5,and is able to satisfy the real security sce-narios of intrusion target detection needs in real security scenarios,and has good generalisation capability.
维普
万方数据