Research And Application Of Human Action Recognition Based On Deep Learning

Human action recognition is an important branch of computer vision and is currently being studied extensively.But traditional human action recognition methods,such as support vector machines,hidden Markovs,and histogram of oriented gradient generally require complex preprocessing of data and its recognition accuracy needs to be improved.With the development of deep learning technology,by establishing an artificial neural network,most complex preprocessing processes can be properly avoided,and the effective features of human movement can be automatically "learned" and the recognition accuracy is relatively high.This paper mainly establishes a deep learning model for the static and dynamic action of the human body.The specific research work is as follows:Firstly,in the static human action recognition,in view of the high complexity of the RGB picture background and the lack of depth information,this paper proposes a M-CNN model to fuse the RGB image information of static human action with the skeletal point information.The skeletal point information effectively compensates for the lack of depth information in RGB image information,reducing the influence of image background complexity on effective feature extraction.The results show the accuracy of static human action recognition is improved after the skeletal point information is integrated.Secondly,in dynamic human action recognition,the current method based on deep learning mostly uses video frames as input.When classifying a long video,it needs to be segmented into short video of equal length to extract features or averages etc.This method not only reduces the accuracy of video classification,but also increases the difficulty of extracting target features due to the complex video background.In order to solve such problems,this paper proposes the DBLSTM-CM model,taking the coordinate sequence of human skeleton points as input,and using DBLSTM to model the time series-based bone coordinate points to extract the main features of dynamic human movement.In order to effectively extract the characteristics of human body action trends,a covariance matrix(CM)is combined with DBLSTM,and the softmax classifier is used for classification.The experimental results show that the accuracy of the DBLSTM-CM model is significantly improved compared with that of the DBLSTM model after adding the covariance matrix.This paper contains 44 figures,6 frames,and 58 references.