LocoESIS: Deep-Learning-Based Leg-Joint Angle Estimation from a Single Pelvis Inertial Sensor

Abstract

Gait analysis is critical for various application areas, including healthcare (assistive robotics), sports, virtual reality, and animation. Generally, an expensive special experimental setup that employs cameras, multiple wearable sensors, or markers has been used for human pose tracking and gait analysis indoors. One of the core aspects of gait analysis or pose tracking is the determination of the joint angle. Therefore, in this study, a neural-network-based joint angle estimation method based solely on pelvis inertial data is proposed and named as LocoESIS (Locomotion Estimator with a Single Inertial Sensor). As a proof of concept, four different neural network models were investigated, including bidirectional long short-term memory (BLSTM), convolutional neural network, wavelet neural network, and unidirectional LSTM. With a lower network complexity, BLSTM resulted in a mean absolute error of up to 5.80°. Moreover, the actual leg motion was compared with the simulated version of the predicted leg joints, confirming that an excellent leg joint estimation result can be obtained even from a single inertial data.