Learning Motion Features from Dynamic Images of Depth Video for Human Action Recognition
Yao Song Huang, Jianyu Yang, Zhanpeng Shao, Youfu Li · 2021 27th International Conference on Mechatronics and Machine Vision in Practice (M2VIP) · 2021
Depth video based action recognition is an important topic in machine vision tasks. Most existing methods design handcrafted features or directly use convolutional neural networks to obtain spatial temporal information in the depth video. However, these methods can not make good use of the 3D information contained in the depth video. In this work, the depth video is projected onto four planes to capture the information of the action in different dimensions. The pixel value of the projection image is determined by the distance between the projected point and the projection plane. Dynamic images of the four projection videos are calculated, which contain different dynamic information of the action on the four planes. Finally, a four-stream network is constructed to learn discriminative motion features. The visualization of the dynamic images shows that the four dynamic images contain different information of the action, and combining them makes it easier for action recognition. The proposed method is evaluated on the benchmark dataset, and the results validate the effectiveness of the method.