Abstract
Human Activity Recognition (HAR) has been an attractive research topic for its applications in areas such as healthcare, smart environments, assisted living, home monitoring, personal fitness assistants etc. Traditional human activity recognition systems are used to recognize common set of activities such as walking, running, sitting, cycling etc. However, these Activity Recognition (AR) systems asks users to provide each activity with large amount of annotations (label) to achieve the acceptable performance. This limitation makes traditional AR systems difficult to extend and to recognize new activities of interest with limited labelled training data. Therefore, it is impractical to assume that users will provide a large amount of annotations since labeling activities is a time-consuming and labor-some process Being able to learn new activities with a limited amount of training data is in demand for practical AR systems. This master project addresses this issue by introducing a novel semi-supervised learning framework for recognizing new activities called SMART, by leveraging the knowledge of the mappings between existing activities with their respective semantic attributes and using limited amount of labelled training data. This model functions in three different spaces. (1) Activity space: recurrent neural network (i.e., LSTM) is induced to transform each activity instance to its semantic attribute representation and also extract the neural embeddings. (2) Semantic Attribute space: based on the activity-attribute knowledge graph, the top-k most likely candidate activities are identified and, (3) Embedding space: density-based clustering (i.e., DBSCAN) is performed on the embeddings of all the instances belonging to the top-k candidate activities. The model is evaluated using the REALDISP activity recognition dataset [1] with 33 physical activities performed by 17 different users. Extensive experiments on real-world data showed that SMART outperformed the state-of-the-art approaches in terms of various metrics for effectively recognizing new/emerging activities.