Abstract
Studying human activities has significant implication in human beneficial applications such as personal healthcare. This research has been facilitated by the development of sensor technologies in pervasive computing with a large quantity of observational data collected about environments and user actions. By mining these data, traditional machine learning techniques have made great progress in recognising activities, but due to the increasing number of sensors and complexity of activities, they are subject to feasibility and scalability. These techniques may benefit from the inclusion of semantic information about the nature and relationships of sensor data and activities being observed. We introduce a new data structure, the context lattice, which allows designers to capture and explore this sort of knowledge. We demonstrate how context lattices can be used to infer human activities with the inclusion of such knowledge. We present comprehensive evaluations of the system against two third-party smart-home data sets, and demonstrate that our approach compares favourably with traditional analytic techniques in many circumstances. We conclude with a discussion of the strengths and weaknesses of context lattices in activity recognition.