Design Of Abnormal Behavior Detection Algorithm Based On Video Surveillance And Implementation Of Embedded System

Prison is an important state organ responsible for detaining and reforming prisoners,and is a vital part of maintaining social stability and tranquility.This article intends to use technology to intelligently transform the prison video surveillance system to assist prison guards,reduce prison supervision workload and improve work efficiency.According to the project of intelligent reformation of a prison in Ningxia in the national key research and development plan undertaken by the school,this paper designs and develops related abnormal behavior detection algorithms and embedded platform system based on deep learning and image processing technology.Finally,the algorithm and system designed and developed in this paper can detect the law of activity and sleep status of prisoners through video surveillance analysis,output visual data that is easy for prison guards to view,promptly warn persons with abnormal behavior,notify prison guards to verify and confirm such personnel,and solve problems in time Take precautions.The algorithm and system designed and developed in this paper are based on the actual needs of the project,with extremely low material and human resources to assist prison guards in the work of troubleshooting,greatly reducing the workload of investigation,greatly improving the efficiency of prison supervision,and early warning of potential prisons through technical means Hidden dangers and excessive behavior.The content of this article is summarized as follows:First of all,most current deep learning-based target detection algorithms have high requirements for computing performance,and are mainly deployed on high-performance desktop computers equipped with GPUs or high-performance servers connected to the network.However,high-performance desktop computers equipped with GPUs are expensive and bulky,making it inconvenient to deploy directly in designated prisons,and at the same time,the cost is too high to exceed the project budget.If data processing is performed by connecting high-performance servers,it will bring hidden dangers to the security of prison data.Aiming at the problem of limited project budget and data security,this paper designs and develops an implementation method of an embedded platform that meets the needs of the project:designing an embedded system based on the NVIDIA JETSON TX2 core board,and building a software environment in which to run a deep learning model to realize the design and development of this paper Algorithm.Experiments show that the embedded system designed based on the TX2 core board is small in size,low in cost,easy to deploy,and the performance in high-power mode can meet the real-time target detection of the Yolov3 algorithm.At the same time,this paper improves the speed of the model detection by pruning the Yolov3 model.Increased by 49.4%,reduced the amount of parameters by 82.3%,and reduced the volume of the model by 82.2%;by using TensorRT to accelerate the Yolov3 inference process,the network model inference speed is increased by about 1.5 times,and the GPU footprint is reduced by 1GB.Secondly,it is difficult for prison guards to find the abnormal behavior problems of prisoners in the daily activities.If humans are used to solve the problem,only the prison guards can observe the records,which requires a lot of manpower investment.Therefore,this paper proposes to use target detection algorithm to realize the target recognition of people and objects in the specified scene,output coordinate information,classification information and system clock information,and then design data visualization algorithm and anomaly judgment algorithm to output intuitive visualization information and anomaly information for prison guards.In this paper,through experimental analysis and comparison of various target detection algorithm principles and performance deployed on embedded platforms,Yolov3 is finally selected as the basic algorithm for system target detection after comparative experiments.The algorithm designed in this paper includes:the improved K-means algorithm automatically obtains the number of cluster centers,so as to realize the automatic clustering of coordinate data and output the hotspot position coordinates and number of hotspots of personnel activities;the visualization algorithm for the length of time people stay in hotspots,after clustering Add clock information to the data to draw and display the distribution of the time and duration of personnel staying at each hot spot;the algorithm of personnel activity heat map drawing,which shows the law of personnel activities in the scene in the form of heat map;the difference value hash algorithm to calculate the activity heat of the day The degree of fit between the graph and the statistical heat activity map.The implementation shows that the improved Yolov3 algorithm in this paper achieves an accuracy rate of more than 99%for the target detection and recognition of the original scene personnel after the self-built data set migration learning training.Due to the strong regularity of the simulated prison scene,the personnel feature changes are small,and the migration learning training There are almost no missed inspections afterwards;data visualization algorithms and anomaly judgment algorithms provide a series of information and anomaly information reminders for prison guards to intuitively output the regularity of the prisoner's activities,which can promptly alert anomalous personnel and greatly reduce the workload of prison guards.Finally,in view of the problem that the psychological activities of prisoners are not easy to be supervised,a program proposes to collect biological information such as the heart rate and blood pressure of the prisoner for an abnormal analysis of mental state through electronic bracelets.The disadvantages of this program are:electronic bracelets are easy to damage and must be replaced frequently Battery or charging,which will bring a lot of follow-up operation and maintenance work and operation and maintenance costs.The project budget is limited,and strive to lower human and material resources into the solution.From a practical point of view,the prison is a special place,and the prisoners will be under video surveillance for 24 hours,so this article designs and develops an image processing algorithm to process the video images of the designated prisoner's sleep state,and outputs the person's sleep state curve,and Various indicators reflecting the sleep status of personnel,based on the obtained curve and indicators to alert persons with abnormal sleep patterns,narrow down the scope of the prisoners who have severe psychological activities,and help prison guards reduce the workload of supervision.According to the experimental surface,this method can effectively detect persons with abnormal sleep state,and with extremely low human and material resources,greatly reduce the workload of prison guards and improve the efficiency of prison supervision.In the system design of this paper,the embedded platform is designed based on the TX2 core board in terms of hardware,which effectively reduces the cost of the hardware system,makes the system deployment more convenient,and satisfies the project requirements well.Software:To build the software environment required for the operation of deep learning models on embedded platforms;use TensorRT to accelerate the algorithm inference process;improve the Yolov3 model and build data sets for migration learning training;design and development of data visualization and anomaly judgment algorithms;sleep state detection Design and development of image processing algorithms.The abnormal behavior detection algorithm and embedded system based on video surveillance designed in this paper successfully meet the requirements of the project and solve practical problems for a prison in Ningxia.