A Neural Network Model to Identify Relative Movements from Wearable Devices
Caroline Sangeetha John, Zachary Mueller, Lakshmi Prayaga, Krishna Devulapalli · 2020
In this paper, we propose a simple neural network model to identify and analyze movement based data collected from wearable devices. The base variables used for the analysis include the relative altitude, the accelerometer acceleration, quaternion, the relative motion of gravity, the user acceleration, and the relative gyro rotation in an x, y, and z plane. Based on previous related studies, extensive feature engineering was utilized to generate 171 unique features from those previously mentioned, which showed a dramatic increase in accuracy. The Label Smoothing Cross Entropy feature was also utilized to make a better generalization of the proposed model. We were able to achieve 95% average on our test data sets which out performed a previous similar work by a good margin of 20%.