Open Set Action Recognition Based on Skeleton

Human action recognition based on skeleton has received increasing attention due to the compact pattern of human skeleton data. Because skeleton data only captures human action information and is not interfered by contextual information such as background and illumination, it is more stable for action recognition. In the real world, we cannot find a dataset that includes all the categories of actions, and there will be many human actions out of dataset, which requires the network to make correct judgment for known actions and reject unknown actions that are not within the scope of cognition. In our work, we propose an Open Set Action Recognition based on Skeleton (OSARS) method to recognition human action in open-set field. We use evidential deep learning to predict category probability and calculate category prediction uncertainty to transform Open Set Action Recognition (OSAR) problem into prediction uncertainty measurement problem.