Research On Human Dangerous Behavior Recognition Method Based On Deep Learning

Human dangerous behavior recognition is an important research field in intelligent monitoring system,which is very critical for crime prevention.Automatic detection of human dangerous behavior in video surveillance scenes in public places such as railway stations,gymnasiums and shopping malls can detect potential accident hazards in time and deal with them immediately,and establish a perfect and immediate early warning mechanism,which can effectively reduce or eliminate public safety problems and ensure personal and property safety.In the real world,the researchers pay more attention to human behavior detection and recognition based on feature attributes,such as video shooting Angle,background brightness,feature diversity and the relationship between human behaviors.To solve the exist problems,the researchers proposed a method to extract information of human behavior skeleton and optical flow characteristic information are extracted from video.At present,the most useful skeletal and optical flow information extraction methods are OpenPose,AlphaPose and OpenCV,etc.Aiming at the problem of human dangerous behavior identification,this work is based on the convolutional network of spatio-temporal graphs and the time-series relation network as training models for detection and recognition of dangerous behaviors of human body.The key of spatio-temporal graphs convolutional network recognition is the extraction of human behavior skeleton.The OpenPose and AlphaPose were used to extract the skeleton sequence of human behavior from the video in the monitoring system,where the skeleton of each frame contains 18 joints of human skeleton and the estimated confidence value of each frame.Then,according to the obtained skeleton feature information,combining with the time vector in the skeleton sequence,a spatio-temporal graphs model was established for risk behavior identification.Finally,the risk level of the behavior is determined by setting the corresponding threshold.Compared with the spatio-temporal graphs convolution network,the time-series relation network uses different feature information.The timeseries relation network establishes a temporal relation network model based on the human body's optical flow information and the temporal relation of video frames.The key of this model is to orderly and randomly extract frames of input time-series relation network from the video.The results from Kinetics and NTU-RGB+D show that both methods can effectively identify multiple dangerous behaviors in different scenarios,and the recognition effect is better than the single feature algorithm with the higher recognition efficiency and the better robustness.Meanwhile,using AlphaPose recognition is more efficient and accurate.The network equipped with time-series relation module can effectively improve the identification accuracy in the detection and recognition of dangerous behaviors.In this paper,a variety of methods were used for comparison experiments,and the results show that the recognition method based on skeleton and optical flow features is better than the algorithm based on manual extraction.