Dependence of Kinect sensors number and position on gestures recognition with Gesture Description Language semantic classifier

We have checked if it is possible to increase effectiveness of standard tracking library (Kinect Software Development Kit) by fusion of body joints gathered from different sensors positioned around the user. The proposed calibration procedure enables integration of skeleton data from set of tracking devices into one skeleton. That procedure eliminates many segmentation and tracking errors. The test set for our methodology was 700 recordings of seven various Okinawa Shorin-ryu Karate techniques performed by black belt instructor. In case when side Kinects were rotated in π/2 and - π/2 around vertical axis relatively to central one number of all not classified Karate techniques dropped by 48% while excessive misclassification error remained in the same level.