Using Smartwatch Data to Estimate Motion

Abstract

In recent years, estimation-based gait analysis has acquired increasing popularity. Current trends in long-term gait monitoring rely on machine learning-based estimations. The potential among different machine learning is examined in this study through a comprehensive analysis of current gait estimates based on those methods. As a result, this research assessed the feasibility of feed forward neural network, long-short term memory, and convolutional neural network longshort term memory for estimating lower limb motion using wrist acceleration acquired from a smartwatch sensor. The wrist acceleration is obtained from a publicly accessible database, which was selected for network training. The gait event is walking, and 32 participants are involved in the gait event. The best result of the lower limb motion acceleration are estimated with correlation coefficient $\gt 0.929$ and root mean square error:$0.36^{\circ} \pm 0.11$% in left ankle acceleration x-axis, $0.22^{\circ} \pm 0.08$% in y-axis, $0.20^{\circ} \pm 0.06$% in z-axis, $0.29^{\circ}\pm 0.11$% in right ankle acceleration x-axis, $0.2^{\circ} \pm 0.09$% in y-axis, and $0.23^{\circ} \pm 0.07$% in z-axis. Besides, the smartwatch sensor method may minimise the requirement for several wearable sensors and provide lower limb estimation on a long-term basis using a simplified sensor design. Consequently, using a smartwatch sensor, the proposed study analyses the viability of employing three different machine-learning models for lower limb estimates based on wrist acceleration.