Research And Deployment Of Video Violence Detection Algorithm For Embedded Devices

Violence is a global challenge,which cannot be avoided in any country or city.Timely and effective detection and suppression of violence is of great significance to protecting people’s lives and property and maintaining social peace and stability.With the continuous promotion of the implementation policy of artificial intelligence algorithm in China’s 13th Five-Year Plan,video-based violent behavior detection has become one of the hot research directions of large-scale implementation of artificial intelligence technology.At present,Video-based violent behavior detection algorithm is mainly used in intelligent content review and intelligent monitoring scenarios,main model of reasoning on the edge of the cloud server or calculation on the box,there is a privacy protection,real-time performance and high deployment cost problem,and the algorithm is transplanted to embedded devices running on can effectively solve these problems,However,in general,the computation power of embedded devices is limited,and some violent behavior detection algorithms based on deep neural networks cannot be deployed in embedded systems due to their high complexity.Therefore,the lightweight of algorithms is particularly important.For embedded devices,this paper studies the lightweight method of video-based violent behavior detection algorithm and carries out embedded deployment of the algorithm.The main work is as follows:(1)To study the effective behavior recognition algorithm TSM,draw lessons from the idea to violence and into the detection task,combining with the characteristics of violence and put forward a kind of violent behavior detection algorithm based on time step into the shift,the algorithm can achieve the performance of 3D convolution with the complexity of 2D convolution,at the same time step shift operation allows the model to detect earlier violence.(2)To study the compression and acceleration technology,the model is further optimized in the design of lightweight convolutional network and model quantization,by introducing the depth of the separable convolution,greatly reduce the number and amount of calculation of the model,through the model of quantitative,32-bit floating point number into INT8 type,improve the operation efficiency of the proposed model.(3)Built a real-time violent behavior detection system based on raspberry PI,and realized the embedded deployment of video-based violent behavior detection algorithm.During the embedded deployment of the algorithm,TVM compiler is used to compile and optimize the model,and TVM Runtime is used to deduce the model on raspberry PI.The realtime frame rate of the system can reach about 30ms,which fully verifies the effectiveness of the proposed algorithm.