Abstract
Stretch sensor-based wearable sensing systems have attracted attention as a technology for human motion tracking in daily life because of their flexibility and wearing comfort. Some previous works proposed methods to estimate joint angles and postures from stretch sensors using machine learning techniques. However, they did not consider sensor shifts due to donning/doffing and long-term use, which may decrease their estimation accuracy. Therefore, this study proposes a joint angle measurement system robust for sensor shifts by using stochastic machine learning. We first developed a smart brace with two stretch sensors instrumented at different heights. For learning an estimation model, the device was gradually shifted and stretch sensor readings of the device and ground truth of knee joint angle from inertial measurement units were obtained during stepping motion. Then, Gaussian mixture models were fitted to the probabilistic distribution of the dataset in each brace shift and the GMMs of all shifts were integrated as the estimation model. By using the model for Gaussian mixture regression, the system estimated joint angle by adaptively changing these models according to brace shifts. The performance of the proposed method was evaluated by comparison with a conventional learning method using measurement data sets at proper and shifted positions.
