Automatic Feature Extraction from Wearable Sensor Data by Use of Machine Learnings

Feasibility of machine leaning techniques for use of automatic detection and analysis of human behavior has been discussed. We focused on the analysis of swing motions in ping-pong play based on the acceleration sensor data of a Smartwatch and to identify slight difference contained in the motion that is unnoticeable by seeing. It has been discussed the feature extraction ability of PCA and auto encoder based on a multilayered neural network combined with clustering methods and shown the possibility of detection of small scale difference in swing motions.