A Correction Method for Monocular Camera-based Gait Joint Angle Estimation Using LSTM With Knee Joint Time-series Information

Gait analysis is a crucial tool for diagnosing neuromuscular disorders and evaluating rehabilitation treatments, but the conventional 3D motion capture systems have limited clinical applicability due to high costs and spatial constraints. This study proposes a novel method for precise lower limb joint angle estimation using monocular camera images and LSTM networks, specifically leveraging knee joint temporal information as the primary input. The system employs HybrIK for 3D pose estimation, uDEAS optimization for initial joint angle calculation, and an LSTM network that predicts hip and ankle movements from knee joint angle sequences by exploiting kinematic coupling between lower limb joints. Validation with 11 participants during overground and treadmill walking showed that the LSTM-based model significantly outperformed the uDEAS-only approach, achieving approximately 40% reduction in estimation errors. Notably, the method enabled reliable ankle angle prediction, which was not feasible with uDEAS alone. By effectively learning inter-joint coordination through knee joint’s temporal patterns, this approach demonstrates that knee joint information can serve as a robust predictor for other joint movements, enabling clinically reliable gait analysis using only a monocular camera.