High-order Features For Action Recognition

Recently,with the continuous development of deep learning,many researchers and employees hope to use deep learning technology to understand videos' content.Action recognition task has attracted the attention of lots of researchers.The recognition of human actions in videos is a classic research topic in the field of computer vision.In practical applications,the task of action recognition is widely used.Many intelligent products take the action recognition algorithm as the core to realize the analysis of video content,such as smart endowment,smart education and so on.Therefore,the action recognition in videos has a broad application prospect and practical significance.In recent years,convolutional neural network has made great progress in the field of computer vision.Because it can automatically extract features without manual selection,this paper uses convolutional neural network to extract high-quality action features in videos efficiently.In this paper,a method of action recognition based on higher-order features is proposed.Firstly,the temporal shift method is used to obtain the spatio-temporal features of the action more efficiently,and bilinear model is used to encode the spatio-temporal features of the action to encode the low-order spatio-temporal features into higher-order features,so as to obtain a higher recognition accuracy.In this paper,a series of experiments are designed to verify the validity of spatio-temporal features extracted by temporal shift method and the promotion effect of bilinear coding on recognition tasks.In this paper,a series of experiments were conducted on the UCF101 dataset.The accuracy of classification was used as the evaluation index to measure the effectiveness of higher-order features on the classification task,and the number of parameters of the method in this paper was listed to describe the method in this paper,which made use of fewer parameters on the premise of ensuring the recognition accuracy.The experimental results show that the temporal shift method can efficiently capture the temporal and spatial features of the action in the video,and the encoded higher-order features can promote the recognition accuracy.