Tracking People across Multiple Non-overlapping RGB-D Sensors

This work presents the development of a surveillance system for monitoring wide area indoor spaces using multiple Kinect devices. The data from these sensors, configured with the widest possible coverage, is integrated into a single coordinate system using a novel calibration technique for non-overlapping range sensors. Moving 3D pixels from each Kinect are transformed into a "plan view" map of activity where the detection and tracking of people is executed. The detection of people is a two step process, data binning and non maxima suppression. The tracking of people is based on the mean-shift algorithm optimized with the prediction step of the Kalman Filter.