Research on aerobics action pose recognition based on deep learning

Taking aerobics as an example, the human movement can be regarded as a series of posture data that changes over time. Compared with other methods, the special kinematic feature model of human skeleton has great advantages in describing the posture change state. In order to achieve the accurate capture of dynamic posture of aerobics, so as to complete the recognition and analysis of motion posture data in a short time, this paper proposes a 3D human dynamic posture recognition method based on Long Short-Term Memory (LSTM) network. First, the first frame model of the 3D human action sequence is selected as the template of the sequence, and the shape difference of the subsequent models of the action sequence is calculated by the shape difference operator relative to the template, which is represented as a low-dimensional shape difference information tensor. Then, the spatial and temporal dimensional features are extracted from the shape difference information tensor by combining two-dimensional convolutional neural network and LSTM to achieve the recognition of human dynamic posture. The above methods were evaluated by the dynamic pose datasets HumanEva, MoSh, SFU, SSM and Transitions; The classification accuracies were 98.4%, 99.7%, 100%, 99.4% and 100%, respectively.