A Contactless Computer Vision System for Underwater Walking and Jogging Gait Analysis Using YOLO-Pose and Multi-CNN BiLSTM Architecture

Abstract

Buoyancy-assisted hydrotherapy exercise has been proved to reduce joint loading and accelerates functional recovery. However, conventional marker or sensor-based approaches are found costly and impractical for underwater use due to water interference and setup constraints for recovery progress monitoring. To overcome these challenges, a computer vision-based gait analysis model was trained for jogging sessions in hydrotherapy pools. This paper used 2D coordinates extracted using You Only Look Once (YOLO) 11m-pose as the model input without noise filtration to validate the its robustness. A comparison of hyperparameter optimization algorithms was also conducted, with the combination of Multivariate Tree-structured Parzen Estimators (MultiTPE) and Hyperband identified as the optimal approach. Two Convolutional Bidirectional Long Short-Term Memory architectures, i.e., single vs. multi convolutional layers (CNNs) per pooling were applied and compared in both multi-head and single-head regression settings. Result found that multi-CNNs per pooling with multi-task learning best exploit inter-parameter correlations. On a 45-samples test set evaluation, the model achieved an Intraclass Correlation Coefficient (ICC) with two-way random effects, absolute agreement, single rater model (ICC(2,1)) of 0.8999, Pearson’s Correlation Coefficient (PCC) of 0.9066, Mean Absolute Error of 0.0954 seconds for swing, stance, and stride time, while 3.5141 steps/min for cadence. The developed system is able to achieve precise analysis for underwater leg movements.