Fitness Movements Recognition and Evaluation Based on LSTM
Yi Wang, Xiaowen Zhu, Chengzhang Qu · Journal of Physics Conference Series · 2019
In this paper, we use Kinect to obtain human skeleton information, then extract 24-dimensional eigenvalues to represent the dynamic movements. A lighted refined LSTM model is employed to recognition the movement serious, finally we propose a movement evaluation model to mainly avoid sports injured. For experiment, we collect 5 kinds of movements as 250 movement samples totally. Using our model, the accuracy can get 80% as the accuracy we test on MSR dataset is 82%. At last, the vulnerable judgment for fitness movement is calculated based on characteristic angles of each dynamic action which can also enlarge the effect of fitness.