Improved Convolutional 3D Networks for Micro-Movements Recognition

Abstract

It is of great significance for computers to recognize the actions in videos. The human body’s action recognition has been applied in many fields. The majority of action recognition methods have relatively low precision in recognizing micro-movements. In some specific scenarios, tasks such as intelligent home companionship for the elderly and early warning for dangerous driving behaviors, the micro-actions of the observed are extremely important in the recognition task. At the same time, due to the physiological characteristics of the elderly or the limitation of the environment, the amplitude of the actions is relatively small. This research suggests an action recognition method based on deep learning to better analyze micro-movements-oriented action recognition. Inspired by transformer, we split an image into fixed-size patches. The network structure of C3D is improved. The idea of image patch is introduced to reduce the receptive field of each region in the video frame. Finally, the experimental verification is performed on two action recognition datasets, UCF101 and NTU. The average accuracies on UCF101 and NTU respectively are 91.74% and 88.01%, which show that the proposed algorithm can effectively improve the recognition ability of micro-movements and obtain better results compared with other baselines.