Public Security Monitoring System Based On Deep Learning

With the continuous progress of science and technology,people’s living standard has also been improved.At the same time,potential threats have gradually infiltrated into people’s life.Many incidents threatening public security occur frequently,which will affect the harmonious and stable development of the society.Due to limitation of the traditional methods and with the rapid development of deep learning,a detection mode based on deep learning network has gradually become popular.Therefore,in view of the statement and analysis of the above problems,in this thesis,a public security monitoring system based on deep learning method is developed.Through this system,targets in the monitoring area can be detected and potential threats can be identified and tracked.The main work of this thesis is summarized as follows:1.Construction and improvement of the deep neural network YOLOv3.Data collection and collation are carried out through the images obtained from the real world and movie scenes,and the data set is made by using the Labelimg tool for image annotation.Then construct three common-used deep learning object detection network models,Faster R-CNN,SSD and YOLOv3 by using Python programming language and Pytorch deep learning framework.Train them on homemade dataset and compare their performances on detection.Through comparison,it is found that YOLO V3 network has relatively high average accuracy with low false detection rate and is more fit on the task of intruder detection.Therefore,YOLO V3 is finally implemented and is optimized from three aspects: training strategy,prior box selection and loss function.Finally,the detection accuracy of the improved YOLO V3 network was improved to 71.62%,and the detection accuracy of Person class,Thugs tools class and Knife class were all improved.2.The realization of potential threat discrimination algorithm and the improvement of tracking algorithm.In order to achieve the goal of threats discrimination,I designed an algorithm which is based on the result of object detection.The algorithm of threats detection is used to judge whether a person is a threat according to the distance relationship and IOU relationship between the detected Person class and other weapons such as Thugs tools and Knife.Then,the virtual fence(protected area)is achieved by using the way of pre-processing the images firstly,then,inputting the processed images to the detection network and finally mapping the results to the original image.For object tracking,I mainly use the algorithm SORT to do it and also improve the SORT by integrating the appearance features to help computing the similarity between two objects.3.Design and Implementation of Public Security Monitoring System.This thesis analyzes the functional and non-functional requirements required by the Public Security Monitoring System,and collates the system logic through the overall design.The software development of the intruder detection system is realized by combining the Py QT5 interface development library and Open CV image processing library.In this thesis,the functional test of intruder detection system is carried out by means of dynamic test and the system operation flow is shown.