In this thesis we give new means for a machine to understand complex and dynamic visual scenes in real time. In particular, we solve the problem of simultaneously reconstructing a certain representation of the world's geometry, the observer's trajectory, and the moving objects' structures and trajectories, with the aid of vision exteroceptive sensors. We proceeded by dividing the problem into three main steps: First, we give a solution to the Simultaneous Localization And Mapping problem (SLAM) for monocular vision that is able to adequately perform in the most ill-conditioned situations: those where the observer approaches the scene in straight line. Second, we incorporate full 3D instantaneous observability by duplicating vision hardware with monocular algorithms. This permits us to avoid some of the inherent drawbacks of classic stereo systems, notably their limited range of 3D observability and the necessity of frequent mechanical calibration. Third, we add detection and tracking of moving objects by making use of this full 3D observability, whose necessity we judge almost inevitable. We choose a sparse, punctual representation of both the world and the moving objects in order to alleviate the computational payload of the image processing algorithms, which are required to extract the necessary geometrical information out of the images. This alleviation is additionally supported by active feature detection and search mechanisms which focus the attention to those image regions with the highest interest. This focusing is achieved by an extensive exploitation of the current knowledge available on the system (all the mapped information), something that we finally highlight to be the ultimate key to success.
展开▼
