Perimeter Intrusion Target Detection System Based On Multi-dimensional Perception

The traditional perimeter security sensing method is single,easily interfered by environmental climate and other factors,and is prone to false alarms,resulting in extremely low detection accuracy of the system.In addition,its system does not have the ability to automatically analyze and intelligently identify,so it can only be used as an auxiliary in modern security system.With the development of cutting-edge technologies such as artificial intelligence,machine vision,and 5G and their applications in smart cities,safe cities and other fields,the intelligent research of perimeter security has received extensive attention.In this paper,a multi-dimensional perception intrusion target detection system is proposed,which can perform two-stage hierarchical detection of intrusion targets in multiple scenarios.Firstly,microwave radar sensor,wireless signal sensor and acceleration sensor are used for target detection,and then the target is accurately detected based on vision.In the detection stage,the three sensors are independent of each other and verify each other,forming a three-dimensional perception field from far to near in the vertical direction of the perimeter,and forming a sensing array in the horizontal direction of the perimeter.For the collected and labeled data,machine learning is used for algorithm analysis.SVM,MLP,and 1D convolution methods are used to conduct experiments and compare the model accuracy and size.In order to realize edge computing,reduce the transmission pressure of the bus and improve the transmission speed of the bus,this paper transplants the model weights obtained by training to the STM32 to run the calculation and optimizes the 485 communication driver module in the Linux kernel of the intelligent substation.In the visual detection stage,the intrusion results obtained in the detection stage are re-examined to further locate and identify the intrusion target.In order to obtain higher detection accuracy,better feature extraction ability and multiscale detection ability similar to YOLOv3 network,this paper improves the two-stage visual detection model Faster RCNN and replaces the backbone network in the network structure,and successively tries to use Res Net50,Res Next50,Efficient Net,Mobile Net backbone network to replace the VGG16 network in the original paper and add the FPN structure as the neck network to obtain multi-scale feature fusion information and multi-feature map output.Finally,in order to extract more critical and important information and enable the model to make more accurate judgments without the model increasing the amount of computation and parameters,this paper also adds the SE channel attention mechanism,CBAM mixed channel and spatial attention mechanism,CA coordinate attention mechanism to the Res Net50 network structure.In this paper,the proposed perimeter intrusion target detection system is designed and implemented,and system analysis and experiments are completed.In the detection stage,the Wiser-MLP and Wiser03-CNN network models are proposed and tested on the self-made dataset,and their accuracies are 86% and 91%,respectively.In the visual detection stage,compared with the original Faster RCNN model,the improved model can increase m AP by16.8% on the Pascal VOC2012 dataset and reached 80.4%.Combining the two-stage detection accuracy,the system’s detection accuracy for intrusion targets is higher than 90%.This paper explores the computing performance of TX2 at the end,and the detection speed of the quantized and accelerated target detection model on TX2 can reach 10 FPS.