Badminton Movement Quality Assessment Through Kinect Sensor Data Using NDT-GCN Algorithm

Human motion quality assessment has broad applications in virtual reality, sports science, and rehabilitation. Current methods for motion assessment predominantly rely on graph convolutional neural networks (GCNs). However, these methods encounter challenges in effectively modeling long-term dependencies, struggling to discern potential connections among nondirectly connected joints in motion, and failing to fully leverage temporal information embedded within frames. To mitigate these limitations, we propose neural ordinary differential time graph convolutional networks (NDT-GCNs). In NDT-GCN, ordinary differential equations (ODEs) are integrated into GCN to extract spatial features from skeletal sequences. We introduce the concept of a complete dependency adjacency matrix, providing a comprehensive depiction of the dynamic relationships among joints. Interframe dependencies are captured through multihead self-attention, while temporal long-range dependencies are extracted using a multiscale convolution module. We establish a dataset comprising human body posture sequences during badminton movements, captured by four Azure Kinect sensors and featuring 12 basic badminton postures performed by 11 participants. Through model comparison experiments and assessments of various influencing factors on this dataset, our results demonstrate that the NDT-GCN model outperforms mainstream models in skeletal feature extraction, achieving an average accuracy of 95.9%. This research introduces a novel approach for modeling and assessing human motion based on skeletal data.