Multi kinect cameras setup for skeleton based action recognition

Abstract

In this research, we investigated and discussed several camera setup methods from different perspectives in order to exhibit advantages and shortcomings, as well as to reveal potentials of other camera setups in terms of image acquisition for action recognition system. Stability and accuracy of tracked skeletons of 10 actions from different views of camera settings are investigated. Experimental results indicates consistency of estimated skeleton lengths regardless of orientation placement of Kinect cameras. Accuracy of estimated skeleton data is less on actions with occluding body parts.