Research On Human Action Recognition Method Based On Deep Learning

The goal of action recognition is to identify the human action in the video.As a basic and challenging task in the field of computer vision,it has broad application prospects in many fields,such as human-computer interaction,virtual reality,intelligent video surveillance and social public security.Traditional action recognition methods rely on manual feature extraction,which has some problems,such as complexity,time-consuming,poor anti-interference and weak generalization ability.In contrast,action recognition method based on deep learning can use neural network model to learn data features autonomously,and it is more efficient and accurate.This paper focuses on the action recognition method based on deep learning.The specific work is as follows:(1)Due to the small sample size of the existing action recognition dataset,overfitting is easy to occur in the model training.In this paper,a data enhancement algorithm based on video is proposed.Under the condition of keeping the data structure and sample label unchanged,this algorithm uses the horizontal translation matrix to cut the video frame image randomly,which makes the dataset sample expanded effectively,so it reduces the risk of model overfitting to a certain extent.(2)Because the data structure of action recognition is video frame sequence,if it is directly input into the model for training,it will lead to huge computational overhead.Therefore,this paper proposes a more efficient video frame sampling algorithm.The algorithm first analyzes the input video frame sequence,and then determines the corresponding sampling strategy.Experimental results show that the algorithm not only solves the problem of computational cost,but also improves the recognition accuracy of the model.(3)For the problem of insufficient spatial feature extraction capabilities in existing deep learning action recognition methods,this paper proposes a human action recognition method with spatial attention mechanism.This method designs a residual network with the convolutional attention module in the process of spatial feature extraction,which enhances the model's ability to extract spatial discriminative features.In addition,in order to solve the deficiency of convolutional attention module in training,this paper also improves the channel attention part,that is,the average pooling features and the maximum pooling features are fused together above all,and then the network weight is trained.Experimental results show that the improved convolutional attention module can make the model more accurate in positioning the key information.(4)Inspired by the spatial attention module,this paper proposes a human action recognition method based on the above research work.This method can not only extract spatial features effectively,but also get better recognition performance by using temporal attention module.Finally,in order to illustrate the rationality of the above research work,experiments are carried out on the open datasets UCF11,KTH,HMDB51 and UCF101.The results show that all the improvement measures can effectively improve the recognition performance of the model.