Research On Violent Behavior Detection Method Based On Deep Learning

"Xueliang project" is a mass public security prevention and control project based on the command platform of the three-level county and rural comprehensive governance centers,supported by comprehensive governance information,based on grid management,and focused on the network application of public security video surveillance.In 2018,it has been written into the No.1 central document for the first time,and has been highly concerned and vigorously promoted by governments at all levels.The construction of multi-level surveillance video networks in cities,counties and villages has begun to take shape.This paper aims at the problem of automatic detection of violence in the project,two key research issues in violent behavior detection are violent behavior identification and violent behavior timing positioning by using deep learning related methods,the application requirements of violence detection in intelligent surveillance video system of "Xueliang project" Software Platform are realized.The following work is mainly completed in this paper.Firstly,aiming at the problem of violent behavior recognition research,this paper proposes a violent behavior recognition method based on the improved R(2+1)D network.Aiming at the problem that the R(2+1)D network identification accuracy is low and the network model parameters are too many,two improvements were made on its network structure:(1)The residual module directly mapping branch structure optimization,which reduces the loss of the space-time characteristics;(2)The residual branches of the residual module are densely spliced with spatio-temporal features,which reduces gradient dispersion.Experiments prove that improved method can effectively improve the recognition accuracy violence.Secondly,aiming at the problem of violent behavior recognition research,this paper proposes a violent behavior recognition method based on the improved R-C3 D network.Aiming at the low average detection accuracy of the R-C3 D network model,two improvements were made:(1)Improve the feature extraction subnet,using R(2 + 1)D alternative C3 D for feature extraction of time and space,which improves the efficiency of spatiotemporal feature extraction;(2)The candidate video clip extraction subnetwork was optimized,and the NMS was optimized as Soft-NMS to screen the candidate video clips,which reduces the missed detection rate of the candidate video clips.After experimental verification method to improve the average detection precision of the violence timing positioning.Finally,this paper developed the violence detection system.The system mainly implements the following functions:(1)The identification function of violence,whether can more accurately identify video clips for violence.(2)The violence behavior detection function,can detect the long video violence in the start time and end time.(3)The local computer and data transmission between the cloud server function,can be trained on the local network model and the cloud data sets are updated.In order to ensure the practicability and stability of the violent behavior detection system,this paper carries out functional tests on each module of the system,and the test results show that the system meets the expected requirements.The research results of this paper can improve the efficiency of automatic detection of video violence in the "Xueliang project",and have a positive role in promoting the maintenance of public security.