Data-driven motion capture using a single smartphone

Abstract

Generating visually appealing human motion sequences using low-dimensional control signals are one big strand of motion research area in computer graphics. We propose a novel approach to reconstruct full body motion controlled by one smartphone. Smartphone is one of the most public devices and includes several inertial sensors such as accelerometer or gyroscope sensor. For correct mapping between full body pose and smartphone sensor, we optimize motion data to low-dimensional coordinate through Gaussian Process Latent Variable Model. Also, our system considers temporal variation by state decomposition model that automatically segments motion phase and connects between latent space and sensor data using nonlinear regression algorithm. Our combined framework allows for reconstructing plausible motion sequences, and we compare the generated animation to ground truth data from commercial motion capture system.