Event-driven body motion analysis for real-time gesture recognition

This paper presents an evaluation of spatiotemporaldata generated by a dynamic stereo vision sensor in ahighdimensional space (3D volume and time) for motion analysisand gesture recognition. In contrast to traditional frame-based(synchronous) stereo cameras, dynamic stereo vision sensorsasynchronously generates events upon scene dynamics. Motionactivities are intrinsically (on-chip) segmented by the sensor, suchthat activity, gesture recognition and tracking can be intuitivelyand efficiently performed. In this work, we investigated theapplicability of this sensor for gesture recognition. We developeda machine lerning method based on the Hidden Markow Modelfor training and automated classifications of gestures using theevent data generated by the sensor. By training eight differentactivities (dance figures) with 15 persons we build up a libraryof 580 recorded activites. An average recognition rate of 97%has been reached.