Research On Human Behavior Recognition In Video

With the rapid development of the Internet and multimedia technology,a large number of videos are shared through the network all the time,and the video information grow exponentially.How to quickly and accurately identify the video content has become a research hot spot currently.Human behavior recognition in videos is an important part of video content recognition,which has a wide application prospect and potential market value in such fields as intelligent video monitoring,human-computer interaction,medical diagnosis,abnormal behavior detection and sports analysis.At the same time,it has also caused profound influence and extensive discussion in the academic circles.Therefore,the study of human behavior recognition in video is of great significance.Aiming at the problems of insufficient feature extraction,redundant video data and poor generalization performance of existing human behavior recognition algorithms,this paper uses computer vision and deep learning related theories and technologies,the research is carried out around the key frame extraction technology and the expression of key features in video,and the human behavior in video is recognized by stages.Firstly,using the advantages of three-dimensional convolutional neural network and channel attention mechanism,the spatiotemporal features are extracted from the decomposed video frames,and the correlation between the channels is modeled.Secondly,using video key frame extraction technology,through the analysis of the difference of human action execution speed and amplitude in each video segment,we can obtain key frames that can effectively represent the video content,reduce the redundancy of data,and improve the characteristics of features.Finally,we use human behavior data set to test the performance of the model,and make an effective judgment of behavior.The main cont ents of this paper are as follows:The human behavior recognition method based on 3DCCA is used to realize the first stage of human behavior recognition.The model is mainly composed of Three Dimensional Convolutional Neural Networks(3DCNN)and Channel At tention(CA).Firstly,RGB video frame sequence is used as the input of the model.Secondly,3D convolutional network is used to extract the spatiotemporal features of the video.Finally,channel attention is used to select the features that are more criti cal to the current behavior recognition from many features.Structural Similarity(SSIM)algorithm was introduced to optimize 3DCCA model to realize the second stage of human behavior recognition.First of all,SSIM algorithm is used to calculate the difference of luminance,contrast and structure between the two frames,and the result is multiplied to attain SSIM value,then select the local and global key frames in the human motion video frame sequence according to the SSIM value,and then selected key frames are used as the input of 3DCCA model to extract behavior features.Finally,the mean value of each feature graph is obtained by global average pooling,and the probability of feature graph is obtained by inputting the value into softmax function to realize human behavior recognition.The experimental results on UCF101 and HMDB51 standard human behavior recognition datasets show that the key frame as the input of the model can effectively improve the rate of behavior recognition.