[2C6-OS-7c] OS-7 (3)

[2C6-OS-7c-03] Action Recognition with Output of Human Pose Estimation

〇Taiki Hirahara¹, Shunichi Taniguchi¹, Atom Sonoda¹, Fujio Toriumi² (1. Lightblue Technology Inc., 2. Univ. of Tokyo)

Keywords:AI, Action Recognition

As the social implementation of AI advances, the demand for estimating what humans are doing from the camera is increasing, and the attempts have been made to estimate the action from the output of the human pose estimation. However, there is a problem that a suitable classifier differs depending on the type of the action, because there are many types of action like motionless action, action taking a short or long time, and having a short or long cycle. It is important to know what classifier is appropriate for the type of action. In this paper, we examined the accuracy changes of the classifier in several actions, focusing on the type of algorithm and the video frames used in the classifier. As a result, we observed that accuracy changes are different between LSTM and SVM, and affected from time interval of video frames or sampled time length.