Human Motion Capture and Recognition Based on Sparse Inertial Sensor

Abstract

The field of human motion capture technology represents an emergent and multifaceted domain that encapsulates various disciplines, including but not limited to computer graphics, ergonomics, and communication technology. A distinct network platform within its domain has been established to ensure the reliability and stability of data transmission. Moreover, a sink node has been configured to facilitate sensor data reception through two distinct channels. Notably, the simplicity of the measurement system is directly proportional to the limited number of sensors used. This study focuses on accurately estimating uncertain human 3D movements via a sparse arrangement of wearable inertial sensors, utilizing only six sensors within the system. The methodology is based on a time series sequence throughout the motion process, wherein a series of discontinuous actions constitute the sequential motion. Deep learning methodologies, specifically recurrent neural networks, were employed to refine the regression parameters. Our approach integrated both historical and present sensor data to forecast future sensor data. These data were amalgamated into a superposed input vector, which was fed back into a shallow neural network to estimate human motion. Our experimental results demonstrate the viability of this approach: the six sensors could accurately replicate representative poses. This finding carries significant implications for advancing and applying wearable devices within the realm of motion capture, offering the potential for widespread adoption and implementation.