Accurate continuous action and gesture recognition method based on skeleton and sliding windows techniques

Vision-based action recognition with the main aim to determine the action class from image sequence has emerged recently thanks to its potential applications. Numerous methods have been proposed for action recognition. However, majority of them mainly focus on isolated action recognition with the assumption that videos are segmented in different clips or shots, each of them contains an image sequence of one sole action. Different from isolated action recognition, continuous action recognition involves recognizing human actions from a continuous stream of input data without requiring pre-segmentation into discrete action instances. In this paper, we propose a method for continuous action recognition. The proposed method incorporates sliding window technique and a light weight action classification model named DDNet. We evaluate the proposed approach on several benchmark datasets. Experimental results show that the proposed method the accuracy of 0.95, 0.97 and 0.71 for IPN, UOW and InHARD datasets respectively on isolated action recognition task. On continuous recognition task, it has achieved a performance of 0.57, 0.76 and 0.33 on IPN, UOW and InHARD datasets.