Human Movement Measurement using Wearable Inertial Sensors and Deep Learning

Abstract

With advances in Micro-Electro-Mechanical Systems (MEMS) technology, inertial sensors have become smaller, cheaper, and are commonly integrated into wearable smart devices, in which their applications have many benefits for our daily lives. They are used in a wide range of applications in human gait analysis, sports training, and healthcare applications. Besides, human movement measurement is an interesting and active topic in sports and medical applications, as it can help people to evaluate physical activities and sports in the daily living environment. Therefore, we focus on the methods for human movement measurement using wearable inertial sensors which are attached to the pedestrian's waist. In this dissertation, the inertial navigation algorithm and deep learning approach for human movement measurement using wearable inertial sensors are proposed to obtain the kinematic and temporal gait parameters. The kinematic parameters, which include the attitude, position, and velocity of the human body, are estimated using the inertial navigation algorithm based on the smoothing algorithm. The temporal gait parameters, such as step length and walking distance, are computed by double integrating acceleration. However, low-cost inertial sensors always have noise and bias that lead to the integration error over time, the algorithm is not possible to estimate the kinematic and temporal gait parameters accurately even for a short distance. Therefore, a known distance straight-line walking trajectory constraint and a constant speed constraint are imposed in the smoothing algorithm. These constraints reduce the accumulation of the integration error even for long walking distances. If the pedestrian walks with a complex walking path or very long distance, the smoothing algorithm needs a very long computation time and large memory requirement to obtain the kinematic and temporal gait parameters, such as the walking trajectory and step length. To tackle this problem, the deep learning-based regression model is proposed to estimate the walking step length and pedestrian traveled distance. The conditional generative adversarial network (CGAN) is used to obtain the step length prediction model, which is trained with a small number of training data. The proposed smoothing algorithm is leveraged to compute the walking step length used as the reference label data in the training stage. Then the step length prediction model is used to estimate walking distance when the testing dataset is obtained. Through the experiments, the proposed smoothing algorithm and deep learning-based step length prediction model can be applied for human movement measurement in real applications with considerable accuracy.