Lateral Walking Gait Recognition and Hip Angle Prediction Using a Dual-Task Learning Framework.

Abstract

Lateral walking exercise is beneficial for the hip abductor enhancement. Accurate gait recognition and continuous hip joint angle prediction are essential for the control of exoskeletons. We propose a dual-task learning framework, the "Twin Brother" model, which fuses convolutional neural network (CNN), long short-term memory (LSTM), neural networks (NNs), and the squeezing-elicited attention mechanism to classify the lateral gait stage and estimate the hip angle from electromyography (EMG) signals. The EMG signals of 6 muscles from 10 subjects during lateral walking were collected. Four gait phases were recognized, and the hip angles of both legs were continuously estimated. The sliding window length of 250 ms and the sliding increment of 3 ms were determined by the requirements of response time and recognition accuracy of the real-time system. We compared the performance of CNN-LSTM, CNN, LSTM, support vector machine, NN, K-nearest neighbor, and the "Twin Brother" models. The "Twin Brother" model achieved a recognition accuracy (mean ± SD) of 98.81% ± 0.14%. The model's predicted root mean square error (RMSE) for the left and right hip angles are 0.9183° ± 0.024° and 1.0511° ± 0.027°, respectively, where the R ² are 0.9853 ± 0.006 and 0.9808 ± 0.008. The accuracy of recognition and estimation are both better than comparative models. For gait phase percentage prediction, RMSE and R ² predicted by the model can reach 0.152° ± 0.014° and 0.986 ± 0.011, respectively. These results demonstrate that the method can be applied to lateral walking gait recognition and hip joint angle prediction.