Dimensionality reduction of data sequences for human activity recognition

Although current human activity recognition can achieve high accuracy rates, data sequences with high-dimensionality are required for a reliable decision to recognize the entire activity. Traditional dimensionality reduction methods do not exploit the local geometry of classification information. In this paper, we introduce the framework of manifold elastic net that encodes the local geometry to find an aligned coordinate system for data representation. The introduced method is efficient because classification error minimization criterion is utilized to directly link the classification error with the selected subspace. In the experimental section, a dataset on human activity recognition is studied from wearable, object, and ambient sensors.