Motion Estimation by Deep Learnings Using Ambient Sensor Network

In this paper, we collect motion data of 7 subjects from an inertial measurement unit and a smartphone simultaneously using our ambient sensor network and then estimate 7 types of human motions of standing up, standing, walking, turning, trotting, sitting down, and sitting by machine learnings. We first perform sensor calibration, coordinate transformation from the sensor coordinate system to the absolute coordinate system, and sensor fusion. Then we compare the accuracy of 4 deep learnings for 3 cases of (1) only accelerations, (2) accelerations and angular velocities, (3) accelerations, angular velocities, and transformed accelerations. As a result, we confirmed that deep convolution long short-term memory achieved 94% accuracy for accelerations, angular velocities, and transformed accelerations.