IEEE Trans Neural Syst Rehabil Eng. 2019 Sep;27(9):1791-1800. doi: 10.1109/TNSRE.2019.2933896. Epub 2019 Aug 8.
In recent years, wearable exoskeletons and powered prosthetics have been considered key elements to remedy mobility loss. One of the main challenges pertaining to this field is the prediction of the wearer's desired motion. In this paper, we perform a human locomotion analysis, and we investigate the accuracy of predicting the angular position of the lower limb joints from the motion of walking canes. Nine healthy subjects took part of this study and performed a locomotor task that comprised straight walking on flat ground, stair ascent, and upright resting posture. Recurrent Neural Networks and polynomial fitting using Least Squares were used to model dynamic and static non-linear mappings, respectively, between the motion of a cane and its contra-lateral leg joints. A successful prediction of both the hip and knee joints was achieved using information from the cane only, and significant improvement of the prediction error was realized through the addition of data from the arm joints. Overall, Recurrent Neural Networks outperformed Least Squares for both joints' angular position prediction. When using the cane only, the static maps were able to predict steady behaviour but failed in predicting transitioning, as opposed to RNN, which was able to capture both steady behaviour and transitions.
近年来，可穿戴式外骨骼和动力假肢被认为是弥补运动能力丧失的关键要素。该领域的主要挑战之一是预测佩戴者的期望运动。在本文中，我们进行了人体运动分析，并研究了从行走手杖的运动预测下肢关节角度位置的准确性。九名健康受试者参与了这项研究，完成了包括平地直走、楼梯上升和直立休息姿势在内的运动任务。使用递归神经网络和最小二乘法的多项式拟合分别用于对手杖运动与其对侧腿部关节之间的动态和静态非线性映射建模。仅使用手杖信息即可成功预测髋关节和膝关节，并且通过添加来自手臂关节的数据，可显著改善预测误差。总体而言，递归神经网络在预测两个关节的角度位置方面均优于最小二乘法。仅使用手杖时，静态图能够预测稳定的行为，但无法预测过渡，而 RNN 则能够捕捉稳定行为和过渡。
