Deep Learning Based Human Movement Classification and Scoring

The objective assessment of human movement quality is essential in domains such as rehabilitation, sports science, and ergonomics, where expert-based evaluations are often subjective, inconsistent, and non-scalable. Recent advancements in pose estimation, notably Google’s MoveNet, facilitate the extraction of 2D skeletal keypoints from video data, enabling automated and objective movement analysis. However, transforming raw keypoints into meaningful movement quality metrics remains a complex challenge due to the need for robust spatial-temporal modeling and feature interpretability. This thesis addresses this challenge by developing a deep learning-based framework that classifies movements as "Good" or "Bad" and predicts expert-assigned quality scores on a continuous scale. The study investigates various architectures—including LSTM, BiLSTM-CNN, CNN, and Temporal Convolutional Networks (TCN)—and integrates feature engineering (joint angles, velocities, body symmetry) and feature selection to enhance model performance. Data augmentation and stratified Group K-Fold cross-validation are applied to ensure model generalization. Experimental results demonstrate that the BiLSTM-CNN architecture achieved the highest classification recall (97.32%) and that regression models reliably predicted movement scores with strong correlation to expert labels. This work contributes a robust, scalable, and interpretable framework for real-time, expert-aligned movement quality assessment.