Research On Violent Behavior Identification Based On Deep Learning

Computer vision is a very important research direction of deep learning.Deep learning can make the computer have a way of thinking similar to human.Now many research topics are how to let the machine replace the human to complete complex and time-consuming tasks.Nowadays,for the use of surveillance video,we need to use deep learning to identify violence,so that the machine can learn to recognize human actions,instead of monitoring the camera screen manually to complete the violence alarm.This paper mainly uses video dataset to train the recognition network.But at present,the influence of complex background interference,light intensity change and other factors on deep learning interferes with the feature extraction effect of neural network.This paper studies these problems as follows:In view of the problem of visual angle gap,illumination change and complex background interference in violent behavior recognition,this paper proposes to use human body contour map as input to train violent behavior recognition network,and human outline image can avoid the influence of background interference and illumination.For the extraction of human outline,this paper uses Deep Lab V3 network for semantic segmentation,extracts the pixels of the human outline in the picture,and generates a binary image.In the gesture recognition experiment using human outline image,the accuracy rate is 83.85%,which proves the feasibility of human outline image in violence behavior recognition.In this paper,Recurrent all-pairs field transforms optical flow algorithm is used to generate optical flow image.In the training of violence behavior recognition network,this paper uses hockey dataset and 3D-CNN to train RGB image,optical flow image and human outline image.The final accuracy is 91.9%,91.0% and 92.5% respectively.In the final model ensemble prediction,the accuracy is 94.75%.The 3D-CNN improves the accuracy of behavior recognition by 8.65% compared with gesture recognition,improves the accuracy of figure contour image by 0.6% and 1.5% compared with RGB image and optical flow image,and improves the accuracy of three input ensemble model by 0.35% compared with double input model.