An improved mobile reinforcement learning for wrong actions detection in aerobics training videos

초록

Currently, mobile network devices are massively popularized. Current mobile reinforcement learning algorithms for motion detection have many problems, such as insufficient fine segmentation and low accuracy. This is mainly because that the primary features without further segmentation of shape-based features cannot recognize the motion well enough. In order to increase the accuracy of wrong actions detection in aerobics training videos, an improved mobile reinforcement learning method is proposed on the intelligent mobile terminals. This method learns and analyzes the preliminary segmentation results of the video image by fractional order network, and refined segments the aerobics training frame by calculating the similarity and coupling force between pixels. Then, motion trajectory feature of the human joint is extracted based on the chunked sampling algorithm in the target region. After extracting the dynamically changing motion trajectory features, the random forest (RF) based dynamic classification model of wrong actions is used to calculate the feature sample confidence by the tree classifier, which finally identifies whether the human motions in the training video are wrong actions. Experimental results show that the human joint motion trajectory is extracted effectively by using the proposed method; and the contrast and uniformity within the region are both greater than 0.9 with segmentation of the aerobics training frames; the wrong actions are dynamically detected in the aerobics training video and the results maintain a high degree of consistency with the proposed method.

키워드