Unsupervised Detection of Motion Primitives in very High Dimensional Sensor Data
Albert Hein, Thomas Kirste · 2010
Abstract. The unsupervised identification of motion primitives in sensor data is widely seen as an important foundation of high-level activity recognition. Currently there are no clustering algorithms capable of processing massive, very highdimensional data sets. In this paper we present an adapted projected stream clustering algorithm, which is able to efficiently detect motion primitives in subspaces of the feature space. The algorithm was tested on> 7GB of motion capturing data from a home care scenario. We evaluated to what extent we were able to detect meaningful clusters using video annotation and real time rendering for visual analysis.