Comparative Evaluation of Classical, Deep, and Quantum Machine Learning for Video-Based Physical Activity Classification

Reliable exercise recognition across schools requires models that remain accurate when acquisition conditions change. This study compares classical machine learning, pose-sequence deep learning, and quantum machine learning to classify four standardized physical-fitness exercises: sit-up, biceps curl, sit-and-reach, and horizontal jump. The original repository contained 2034 videos. A SHA-256 audit of the temporal pose arrays identified 152 exact duplicate copies, all from the same school, leaving 1882 unique videos for the primary analysis. We used YOLO26n-Pose to extract 17 COCO keypoints. We evaluated two input representations: 110 aggregated biomechanical descriptors and 128-step sequences containing 34 local pose coordinates. Model selection and evaluation followed a nested Leave-One-School-Out protocol. Random Forest obtained the highest pooled outer-fold performance (Macro-F1 = 0.9964, accuracy = 0.9968, MCC = 0.9953), with six errors in 1882 videos; XGBoost and RBF-SVC followed with Macro-F1 values of 0.9932 and 0.9881. Among the deep models, BiLSTM, CNN1D, and ST-GCN achieved Macro-F1 values of 0.9711, 0.9666, and 0.9498, respectively. A paired stratified bootstrap supported the Random Forest advantage over XGBoost for Macro-F1, although the Holm-adjusted exact McNemar test for overall correctness was not significant. In exploratory controlled QML experiments, neither QSVC nor VQC exceeded its matched classical baseline under the tested configurations. These results show that compact biomechanical descriptors provide a strong and stable representation for cross-school exercise recognition in this dataset.