Non-Contact Fatigue Estimation in Healthy Individuals Using Azure Kinect: Contribution of Multiple Kinematic Features

Monitoring exercise-induced fatigue is important for maintaining the effectiveness of training and preventing injury. We evaluated a non-contact approach that estimates perceived fatigue from full-body kinematics captured by an Azure Kinect depth camera. Ten healthy young adults repeatedly performed simple, reproducible whole-body movements, and 3D skeletal coordinates from 32 joints were recorded. After smoothing, 24 kinematic features (joint angles, angular velocities, and cycle timing) were extracted. Fatigue labels (Low, Medium, and High) were obtained using the Borg CR10 scale at 30-s intervals. A random forest classifier was trained and evaluated with leave-one-subject-out cross-validation, and class imbalance was addressed by comparing no correction, class weighting, and random oversampling within the training folds. The model discriminated fatigue levels with high performance (overall accuracy 86%; macro ROC AUC 0.98 (LOSO point estimate) under oversampling), and feature importance analysis indicated distributed contributions across feature categories. These results suggest that simple camera-based kinematic analysis can feasibly estimate perceived fatigue during basic movements. Future work will expand the cohort, diversify tasks, and integrate physiological signals to improve generalization and provide segment-level interpretability.