ESPar: Mining emerging sequential patterns for activity recognition in body sensor networks

Wireless Body Sensor Networks offer many promising applications in healthcare, well-being and entertainment. One of the emerging applications is recognizing activities of daily living. This task is particularly challenging because in real life people often preform activities in not only a simple (i.e., sequential), but also complex (i.e., interleaved and concurrent) manner. Existing solutions typically require proper training for building the models for interleaved and concurrent activities, hence they may not be exible to real-life deployment. In this paper, we build a wireless body sensor network to investigate this challenging problem, and introduce a knowledge pattern named Emerging Sequential Pattern (ESP) - a sequential pattern that discovers the signi cant differences - between activity classes. Leveraging on ESPs, we build our complex activity models directly upon the sequential model, eliminating the training process. We conduct a real-world trace collection using our wireless body sensor network in a smart home, and conduct comprehensive empirical studies to evaluate and compare our solution with the state-of-the-art solutions. The results demonstrate that our system achieves an overall accuracy of 91.89% for recognizing sequential, interleaved and concurrent activities, outperforming existing solutions.