EXPERIMENTS ON DEEP LEARNING FOR WEARABLE ACTIVITY RECOGNITION

Abstract

Current research is beginning to adopt deep neural network models on human activity recognition to extract features automatically from sensor data rather than relying on carefully designing suitable feature representation. However, there is just a few of custom deep architectures are explored. In this paper, we presents experiments on three deep leaning models for human activity recognition using wearable sensor. The effectiveness of the three deep neural networks is validated on accelerometer data from two public datasets. The results show that with enough sensor input data, highway convolutional networks provide higher accuracy than the other deep learning models.