Abstract
The aim of activity recognition is to identify automatically what a person is doing at a given point in time from a series of observations. Activity recognition is a very active topic and is considered an essential step towards the design of many advanced systems. Recently, mobile and embedded systems have received growing interest as context-sensing platforms for activity recognition. However, these devices have limited battery life and do not allow continuous user tracking. In this paper, we present a novel activity tracking method integrating a dynamic programming algorithm for sequence alignment into a nearest-neighbour classifier. Our scheme is capable of filling gaps in sensed data by exploiting long-range dependencies in human behaviour. Initial experiments on a standard dataset show very promising results even with little training data.