Research And Implementation Of Human Behavior Recognition Technology Based On Surveillance Video Stream

In recent years,video identification is a difficult and hot topic in the field of computer vision.At present,most researchers mainly focus on human behavior in static pictures.The method of static pictures ignores the continuity of human behavior,and there is a lack of time correlation between samples.Considering the time sequence of human behavior,this paper uses video stream as data source to analyze the human motion characteristics.Starting from the research of walking,standing,sitting and falling in a certain scene,it realizes the classification and recognition of different human behaviors in video stream or image sequence.The main research work of this paper is as follows:(1)The human behavior recognition methods based on surveillance video streams at home and abroad are compared and analyzed.The related behavior recognition technologies are studied,and the advantages and disadvantages of various algorithms are analyzed.On this basis,the research methods of this paper are designed and new ideas to solve the problems are put forward.(2)Using the clipped video stream with single label category as data source,the improved dense trajectory algorithm is applied to extract the trajectory features,HOG features,HOF features and MBH features of human behavior,and the Fisher Vector coding processing of various features is realized,and the L2 norm normalization processing is carried out.Combining the temporal features extracted by three-dimensional convolution dynamic neural network,the optimized feature fusion algorithm which assigns certain weight coefficients and different kernels to various features makes the behavior features more specific than the single feature.The supervised learning method is used to train the human behavior recognition model in the collated behavior data samples,the human behavior recognition model is tested on the public data set HDMB51 and My-database respectively,the tested samples are output the behavior label and accuracy.(3)In the experimental analysis,the data samples come from a variety of open data sets,and the sample data collected in daily life are added.The accuracy of human behavior recognition in different scenarios is tested with the trained model.The experimental results show that the proposed method is robust to disturbances such as image contrast change,brightness change and noise,This method not only improves the accuracy of behavior recognition,but also applies to other human behavior recognition.