Research And Implementation Of Violence Detection Algorithm Based On Surveillance Video

With the development of new media and the acceleration of information dissemination,when information enriches daily life,negative information also follows.People are more likely to be exposed to violent,pornographic and bloody videos.The frequency of violent incidents is higher and higher,which has caused bad effects.With the improvement and full coverage of surveillance systems,physical violence detection has become a research hotspot in the area of human activity recognition.Traditional surveillance systems cannot obtain information from video.The extraction of information depends on people.Violence occurs for a short time but causes great harm.Therefore,it is necessary to detect and analyze human physical violence in real time.In order to stop and reduce the harm caused by violence in time.This paper proposes a physical violence detecting method based on surveillance video.Each camera captures video images and extracts bone models.First extract features from bone data.Then use the dataset to train the classifier.Finally,the model is transplanted to embedded devices.This article establishes the violent detection system based on Raspberry Pi.Once a violence event is detected,the system will send an alarm.It makes the monitoring system more intelligent.This paper proposes an optimized bone model extraction method for multi-person recognition.First simplify the bone model.The COCO skeleton model removes unnecessary bone points on the head to form an improved COCO skeleton model.Then,the OpenPose model is optimized for the problems of high redundancy and poor real-time performance.In the feature processing stage,the Mobile Net network is used to replace the structure of VGG-19.In the initialization phase and the refinement phase,the parallel convolution is changed to a convolution structure by sharing the convolution structure.This method reduces the redundancy of the model.The optimized model training speed is faster than the original model.The trained model is reduced from 203.9MB to 7.8MB,which can effectively reduce hardware costs.By comparing the detection time of images containing a single person,two persons,and four persons,the detection time of the improved model is reduced by 22.37%,23.04%,and 19.30%respectively compared with the original model.Therefore,the real-time performance of the optimized model is higher than original model.Then,features are extracted and selected based on the location data of the bone points.First,we preprocess the bone data to normalize the size of the frame image.Aiming at the problem of missing bone points caused by occlusion,this paper proposes a key point filling algorithm to improve the bone models.We get a complete human skeleton model.By analyzing the skeletal point data of violent actions,a total of 85 groups of morphological and dynamic features are extracted for a sliding window.The de-redundancy Relief-F algorithm is proposed to select features.Principal component analysis is used to reduce the dimensionality of features.Finally,a 26-dimensional feature set is obtained.Then we use the support vector machine algorithm to classify the actions.The support vector machine algorithm uses the RBF kernel function with the best classification result.Use the grid search method to find the optimal parameter combination.The final classification result is obtained using five-fold cross-validation:the accuracy of violent action recognition is 97.4%.Finally,this paper designs a violent motion detection system based on surveillance video.First transplant the improved bone extraction algorithm to the Raspberry Pi.The system uses the camera mounted on the Raspberry Pi to collect video data.Extract the skeletal model of the human target from the video frame.Then use the trained violent action recognition model to classify the actions.Finally,the recognition results are visualized in the monitoring center.Once a violence event is detected,the system will send an alarm.