Towards Kinect-based wheelchair user behavior analysis

In this thesis we present some techniques which can help to estimate the level of attention of a person sitting on a wheelchair in real time. We focus on person detection, head pose estimation and facial features localization, using depth data given by the Kinect, a low-quality consumer depth camera. Our solution to detect a person in the image is based on the flood fill algorithm and requires no initialization. We estimate the head pose adopting the discriminative random regression forests approach and we filter the solution in time domain, achieving robust results. Finally, we try to extract the 3D locations of a set of chosen facial points using random regression forests, which have been trained from a dataset acquired with the Kinect sensor.