Research on Deep Learning Based Badminton Players' Movement Analysis Methods

This research utilizes human posture estimating technology to feather motion capture to overcome the issues of low precision of motion evaluation and inadequate real-time performance in classic motion capture analysis approaches. The evaluation of standard degree and the identification of badminton swing are examined in badminton instruction. First, a simplified OpenPose-based lightweight human posture estimation model is put forth: The 7x7 convolutional kernel structure inside the model is modified to a 7x7 convolutional kernel structure, and the lightweight MobileNet network replaces the Visual Geometry Group19(VGG19) feature extraction network. The model was constructed as a tandem structure with a 1×1 convolution, a 3×3 depth-separable convolution, and a null convolution with an expansion coefficient of 2. Test results indicate that the OpenPose model, which is lightweight, increases performance threefold and increases the accuracy of bone point recognition on the human arm by 3.57%.