Research And Application Of Violence Detection System Based On Deep Learning

In recent years,the development and application of deep learning technology has greatly promoted the rapid development of a series of related fields such as computer vision,natural speech processing,and artificial intelligence.As an important branch and research hot spot in the field of computer vision,abnormal behavior detection and analysis are closely related to people's lives.Violence detection refers to the timely and effective detection of violent behavior in video,and is an important part of abnormal behavior detection.The detection of violence based on surveillance video is the basis for protecting the personal safety of the people and maintaining the peace and stability of the social environment.It is also an important part of modern intelligent video surveillance and has important research significance and application value.The content of this paper is based on deep learning violence detection.The current excellent slow fast idea is used to design a three-dimensional convolutional neural network that matches the capacity and complexity of the task,make a data set for violence detection tasks,and train the designed neural network on the above data set to obtain the violence detection task Parameter model.And use neural network and model combination to complete the task of violent behavior detection.Finally,two online and offline detection methods are designed and implemented.The problem of online detection in real time is solved by multi-threaded concurrent methods.The main work of this paper includes the following parts:(1)The design of neural network structureIn this paper,a three-dimensional convolutional neural network with fast and slow channel is designed to extract the spatiotemporal information of video.The three-dimensional convolution operation can capture the relationship between time and space in the video.The fast and slow channels can extract spatial information and timing information in the video in a targeted manner.The combination of three-dimensional convolution and network with fast and slow channel includes the advantages of both.Experiments show that the network has a clear hierarchy and a reasonable structure,and achieves fast classification while ensuring accurate classification.(2)Production of violence data setThe application scenario of the algorithm in this paper is in the field of security monitoring.At present,there is no clear,fixed angle,continuous image violence data set.In this paper,a violence data set for monitoring is produced for training the model.The data set contains five common acts of violence:punching,kicking,strangling the neck,wrestling,and holding arms.A total of about 1600 video samples are included,including multiple image acquisition angles,and all angles are between 45-60 degrees in combination with the actual use scene.(3)Training of violent behavior parameter modelIn order to improve the generalization ability of the network model,this paper has determined the best training strategy after many comparative experiments.The NAG algorithm was chosen as the optimization algorithm because it is very flexible.The optimal initial learning rate used in this paper is 0.05,and use cosine function decay as learning rate decay strategy.In order to prevent overfitting,the network model introduces 12 regularization and Dropout.(4)Implementation of violence detection system.This paper has designed a violence detection system based on deep learning,which can effectively avoid artificial missed inspection due to visual fatigue or inconcentration,improve detection efficiency,save manpower and material resources.The system designed in this paper is divided into two modes:offline detection and online detection according to the input data.In the offline detection mode,the input is a recorded surveillance video.The system judges whether violence occurs in the entire video and locates the time when the violence occurred.In this mode,the system pays more attention to accuracy.In the online detection mode,the input is real-time video data to detect violence in real time.In this mode,the system pays more attention to real-time.The experimental results show that the method studied and implemented in this paper can effectively detect the violence in the video.The real-time performance and detection rate of the algorithm can meet the actual needs.