Character action recognition based on deep convolutional neural network and action sequence
Jiaen Li, Fuji Ren, Shun Nishide, Xin Kang · 2019
Object motion recognition has been widely used in video tracking, motion analysis, medical assistance and human-computer intelligent interaction, etc., which has become a hot research field in the field of computer vision. One of the main difficulties in character action recognition is to extract powerful features from video input to describe character action. It is not only necessary to detect the characters from the complex background, but also to accurately identify the changes between the movements [1]. In the process of character action research, the accurate description of human skeleton points makes the research on the key points have high accuracy and real-time. This paper proposes a motion recognition method based on depth feature and motion sequence, and adds depth convolution information on the basis of motion unit. The powerful feature extraction and description ability of convolutional network are employed to further describe and analyze the action objects[2]. The Transfer learning and structure optimization of the deep convolutional network were carried out, and the training and data testing were carried out on the new data set, so as to improve the recognition rate of human actions.