Unsupervised Representation Learning Method In Sensor Based Human Activity Recognition

Deep learning methods contribute to improve the estimation accuracy in human activity recognition (HAR) using sensor data. In general, the dataset used in HAR consists of accelerometer data and activity labels. Because of the widespread use of mobile devices, large amount of accelerometer sensor data without activity labels can be easily collected. The problem of annotation needs a large amount of time-consuming cost and human labor to annotate a activity labels to recorded sensor data. Therefore, we need a method to make deep learning models acquire feature representations from accelerometer data without activity labels in HAR. In this study, based on the unsupervised representation learning method proposed in image recognition, we proposed a new unsupervised representation learning method which combines segment discrimination (SD), autoencoder (AE) and feature independent softmax (FIS). Our experimental results showed that our proposed method outperformed the conventional method in fine-tuning accuracy in HAR.