Residual Neural Network and Long Short-Term Memory–Based Algorithm for Estimating the Motion Trajectory of Inertial Measurement Units

Abstract

Motion trajectory is a technology used in smart medicine, rehabilitation engineering, and interactive games. Studies have proposed several motion trajectory tracking systems based on inertial measurement units (IMUs). However, these systems are beset by problems such as low adaptability to various activities and the capacity for only two-dimensional trajectory estimation. To solve these problems, this study proposed a three-dimensional (3D) motion trajectory tracking system based on a residual neural network (ResNet) and bidirectional long short-term memory (Bi-LSTM) to estimate the user’s trajectory of daily activities. A participant was asked to wear the IMU device and perform two types of specified activities, namely hand-movement tasks and walking. Nine degrees of freedom data from the IMUs were collected during these activities and input into the deep learning model, which combines ResNet and Bi-LSTM, to detect features and predict the motion trajectory. Several experiments were conducted to obtain the optimal system architecture and parameter settings and to evaluate the accuracy of motion trajectory estimation. The accuracy of the proposed system was also compared with those in other studies. The results demonstrated that the relative trajectory error of walking is 0.81 m and that of general cases is 0.54 m, which all outperforms the other related studies. The evidence indicated that the proposed system can predict the 3D motion trajectory of daily activities accurately and can be easily applied to track the motion of most daily activities.