Transitional Activity Recognition with Manifold Embedding

Abstract

Activity monitoring is an important part of pervasive sensing, particularly for assessing activities of daily living for elderly patients and those with chronic diseases. Previous studies have mainly focused on binary transitions between activities, but have overlooked detailed transitional patterns. For patient studies, this transition period can be prolonged and may be indicative of the progression of disease. To observe, as well as quantify, transitional activities, a manifold embedding approach is proposed in this paper. The method uses a spectral graph partitioning and transition labelling approach for identifying principal and transitional activity patterns. The practical value of the work is demonstrated through laboratory experiments for identifying specific transitions and detecting simulated motion impairment.