Deep Learning-Based Quantitative Evaluation Model for Student Motor Skills and its Application Validation

Abstract

Quantitative assessment of student motor skills requires robust spatiotemporal feature extraction from motion video data, especially when automated evaluation is expected to reduce subjective scoring bias. This study develops a deep learning-based evaluation model integrating pose recognition, convolutional neural networks, and long short-term memory networks. Movement videos are preprocessed to extract key human-joint points, CNN layers are used to capture spatial posture features, and LSTM layers model temporal continuity and coordination during movement execution. A multi-task regression module is further introduced to quantify posture stability, movement fluency, and body coordination. The model is validated using movement data from 1200 students covering basketball, gymnastics, and long jump. Experimental results show that the mean squared error of the scoring task is 0.042, with repeated-test variation controlled within 0.002, indicating high consistency and generalization across different movement types. The proposed framework provides an engineering-oriented solution for intelligent motion assessment, video-based sensing, and automated spatiotemporal signal analysis.