LEARNING EMBEDDINGS FOR WEARABLE-BASED HUMAN ACTIVITY ANALYSIS

Abstract

The embedded sensors in widely used smartphones, wearable devices and smart environments make the sensor data stream of human activity more accessible. With the development of deep neural networks, extensive studies have been conducted using deep learning methods to extract useful information from the sensor data to recognize the human activity, identify the person, or monitor the health condition of the person. However, applying deep neural networks to the sensor based human activity analysis task remains a challenging research problem in ubiquitous computing. Some of the reasons are: (i) The majority of the acquired data has no labels; (ii) Most of the previous works in activity and sensor stream analysis have been focusing on one aspect of the data, e.g. only recognizing the type of the activity or only identifying the person who performed the activity; (iii) Segmenting a continuous sensor stream and preserving the completeness of each human activity is difficult. In this dissertation, various deep learning techniques have been studied to address these problems in a weakly supervised, unsupervised, or semi-supervised manner. All the developed techniques use deep learning networks to learn embedding spaces in which activities group and thus classifiers can be trained efficiently. For this, both siamese network architectures for weakly supervised data and autoencoder type networks for unsupervised techniques are learned and combined.