Research On Human Abnormal Behavior Recognition Based On Deep Learning

With the continuous rapid development and widespread use of technology and deep learning technology,as well as the rapid rise of computer vision technology and widespread attention,human abnormal behavior recognition technology,as an important branch of computer vision,naturally attracts much attention.At the same time,human abnormal behavior recognition technology is widely used in security,human-computer interaction,safe driving,video retrieval and other fields.In recent years,some automatic recognition algorithms based on deep learning technology have also been applied to the field of behavior recognition.Compared with the traditional neural network,the convolutional neural network,which is representative of deep learning,improves the efficiency of behavior recognition and has achieved breakthrough success in other fields of computer vision.Therefore,it is of great research significance to use deep learning to efficiently and accurately identify abnormal behaviors that appear in videos.In the study of video,there are relatively few studies on the recognition of abnormal human behavior in some specific environments.Due to the diversity,randomness and unpredictability of human abnormal behavior,the research on abnormal behavior has not achieved good results.Therefore,this paper based on deep learning to conduct research on human abnormal behavior recognition,and achieved good recognition results.The main contents of this article are as follows.First,it introduces the theoretical knowledge of deep learning,and studies some common traditional algorithms of human behavior recognition and the emerging deep learning algorithms.The optical flow algorithm and the dual-flow convolutional neural network model are focused on.In the dual-stream convolutional neural network,it is composed of a spatial stream network and a temporal stream network.They have the same network structure and depth.The difference is that the spatial flow network uses RGB images as the network input,while the time flow network uses the optical flow graph calculated by the optical flow method as the network input.Therefore,the dualstream convolutional neural network can make full use of the spatial and temporal information features of the video,and improve more available information features for human abnormal behavior recognition.Second,in order to meet the needs of human abnormal behavior recognition in a specific environment,it is necessary to analyze the behaviors displayed by characters in the experimental environment,and then define some of them as abnormal behaviors.Because the environment of this experiment is a graduate studio,the characters fall(fall),collide(collide),hug(hug),shake(handshake),fight(fight),sleep(sleep),eat(eat)This behavior is defined as abnormal behavior,and as the behavior category to be identified in this topic,other behavior categories are not studied.Third,human behavior is composed of continuous movements over a period of time.Recognition of human behavior must include not only the contour features of the movement in space,but also the characteristics of the movement trend of the movement in time,so double-stream convolution The neural network model is the most reasonable.At the same time,this paper proposes an improved dual-flow network,which uses the VGG16 network instead of the convolutional neural network in the dual-flow convolutional neural network.Because there are not many behavior categories to be identified in this paper,some parameters of VGG16 have been modified.For the features of the improved dual stream network extracted from the spatial stream network and the temporal stream network,a weighted fusion method is used for fusion experiments.The experimental results under different weights are compared to obtain the weight with the highest recognition rate.Fourth,in order to improve the recognition rate of human anomalous behavior,the optical flow diagram obtained by optical flow method is superimposed on 2 frames,5 frames,and 8 frames,and then experiments are performed by superimposed optical flow diagrams.It is verified that the superimposed optical flow diagram contains more information about the movement trends of people,and the recognition rate of abnormal behavior recognition has been improved to a certain extent,and good recognition results have been obtained.