Wide area motion capture using an array of consumer grade structured light depthsensors

In this thesis we propose a solution to how a system can find and track people, as well as recognizing their gestures, in a 360◦ field of view using consumer grade products. We describe a system connecting multiple depth cameras in an array and have them operate as a single camera controlled by a single computer. Using a single camera providing features such as detection, tracking and recognizing gestures of people, we specifically focus on the difficulties of preserving these features in moving forward to an array of cameras. We propose a solution based on Microsoft Kinect and Kinect SDK, using linear transformation to account for a fixed camera model to combine skeleton data from an array of Kinect sensors. Furthermore, we use positional based identification to determine whether people are being tracked by another camera in the system. The contributions of this work include insight into the challenges of building this kind of system based on Kinect hardware and software intended for use on a single computer, such as performance bottlenecks, along with possible alternative solutions. In particular, we present performance measurements for a single computer running up to four sensors and show a system that can run satisfactorily with up to at least 5 sensors on today’s computers. We show what requirements on hardware can be expected for such a system, as well as where there are potential limits as the number of sensors increase.