User interest profiling by real time person detection and coarse gaze estimation

This thesis makes different contributions to person detection, coarse gaze estimation and user interest profiling. We have proposed two methods to reduce the complexity of a multi-scale person detection, which address the two fundamental bottlenecks of cascade detectors: the number of weak classifiers that need to be evaluated in each cascade, and the total number of detection windows that must be evaluated. As regards the task of people profiling, we proposed a strategy to profile the attention of people moving in a known space, exploiting coarse gaze estimation and a novel model based on optical flow to improve attention prediction, without the need of a tracker.