Multi-Inertial Sensor-Based Arm 3D Motion Tracking Using Elman Neural Network

Abstract

Recent years have witnessed the rapid development of microelectromechanical systems, and human motion tracking technology based on IMU (inertial measurement unit) has attracted much attention. However, the magnetic field varies with time and position, which makes it necessary to calibrate sensors before tracking. To address the poor adaptability of IMU to the environments and improve the accuracy of estimated traces, this paper presents an ENN-based (Elman neural network) method to track human arm motions, which consists of two steps. First, the data derived from IMUs are preprocessed for the rough Euler angles; then, an ENN is trained to estimate motions. We explore the initially estimated position to calibrate the acceleration measurements as the input of the ENN. Real-world experiments of arm motion tracking are carried out with the ground truth from an optical motion tracking system. The experimental results show that the mean tracking errors are around 35 mm, with a strong ability to eliminate the effect of extreme measurement and environment noises, avoiding calibrating the magnetometer. The implementation of the well-trained model to independent motions indicates that the robustness of the proposed method is excellent, and the errors reduce by 37.2% on the $x$ -axis and perform similarly on the $z$ -axis compared with 4 traditional methods. This method quite suits those situations where trajectory tracking of the standardized motions is required, such as the medical habilitation.