People Detection and Tracking with Kinect for Mobile Platforms

Human detection is a key ability for robot applications that operate in environments where people are present, or in situation where those applications are requested to interact with them. It’s the case for social robots like aids for the rehabilitation of inmates in hospitals, assistance in office, guides for museum tours. In this thesis we will investigate on how we can make use of the new Microsoft’s gaming sensor, the Kinect, to address the issues of real-time people detection and tracking, since the sensor has been built in order to detect people and track their movements. We developed a system that is able of detecting and tracking people in near real-time both on fixed environments and mobile platforms. We tested four different classifiers on different situations. The best classifier showed very good detection and tracking results whereas, because of some segmentation problems, the performances of the complete system have been subjected to a lowering with respect to the theoretical ones. We developed also a method for getting rid of some of these segmentation problems and it showed some improvements for the complete system together with some drawbacks that affected the theoretical results. However the complete system works good and with a frame rate of 2 fps on average. Most of the computational load is due again to the segmentation module, so an improvement of this module would lead to both improvements on the real-time performances and on the detection results