This thesis focuses on the various aspects of autonomous environment learning for indoor service robots. Particularly, on landmark extraction from sensor data, autonomous map building, and robot localization.

To univocally identify landmarks from sensor data, we study several landmark representations, and the mathematical foundation necessary to extract the features that build them from images and laser range data. The features extracted from just one sensor may not suce in the invariant characterization of landmarks and objects, pushing for the combination of information from multiple sources. We present a new algorithm that fuses complementary information from two low level vision modules into coherent object models that can be tracked and learned in a mobile robotics context. Illumination conditions and occlusions are the most prominent artifacts that hinder data association in computer vision. By using photogrammetric and geometric constraints we restrict the search for landmark matches in successive images, and by locking our interest in one or a set of landmarks in the scene, we track those landmarks along successive frames, reducing considerably the data association problem. We concentrate on those tools from the geometry of multiple views that are relevant to the computation of initial landmark location estimates for coarse motion recovery; a desirable characteristic when odometry is not available or is highly unreliable.

Once landmarks are accurately extracted and identied, the second part of the problem is to use these observations for the localization of the robot, as well as the refinement of the landmark location estimates. We consider robot motion and sensor observations as stochastic processes, and treat the problem from an estimation theoretic point of view, dealing with noise by using probabilistic methods.

The main drawback we encounter is that current estimation techniques have been devised for static environments, and that they lack robustness in more realistic situations. To aid in those situations in which landmark observations might not be consistent in time, we propose a new set of temporal landmark quality functions, and show how by incorporating these functions in the data association tests, the overall estimation-theoretic approach to map building and localization is improved. The basic idea consists on using the history of data association mismatches for the computation of the likelihood of future data association, together with the spatial compatibility tests already available.

Special attention is paid in that the removal of spurious landmarks from the map does not violate the basic convergence properties of the localization and map building algorithms already described in the literature; namely, asymptotic convergence and full correlation.

The thesis also gives an in depth analysis of the fully correlated model to localization and map building from a control systems theory point of view. Considering the fact that the Kalman .lter is nothing else but an optimal observer, we analyze the implications of having a state vector that is being revised by fully correlated noise measurements. We end up revealing theoretically and with experiments the strong limitations of using a fully correlated noise driven estimation theoretic approach to map building and localization in relation to the total number of landmarks used.

Partial observability hinders full reconstructibility of the state space, making the .nal map estimate dependant on the initial observations, and does not guarantee convergence to a positive de nite covariance matrix. Partial controllability on the other hand, makes the .lter beleive after a number of iterations, that it has accurate estimates of the landmark states, with their corresponding Kalman gains converging to zero. That is, after a few steps, innovations are useless. We show how to palliate the e.ects of full correlation and partial controllability. Furthermore, given that the Kalman .lter is an optimal observer for the reconstruction of fully correlated states; it seems pertinent to build an optimal regulator in order to keep the robot as close as possible to a desired motion path when building a map. We show also how the duality between observability and controllability can be exploited in designing such an optimal regulator.

Any map building and localization algorithm for mobile robotics that is to work in real time must be able to relate observations and model matches in an expeditious way. Some of the landmark compatibility tests are computationally expensive, and their application has to be carefully designed. We touch upon the time complexity issues of the various landmark compatibility tests used, and also on the desirable properties of our chosen map data structure.
Furthermore, we propose a series of tasks that must be handled when dealing with landmark data association. From model compatibility tests, to search space reduction and hypothesis formation, to the actual association of observations and models.

The work presented in this thesis spans several areas of engineering and computer science, from new computer vision algorithms, to novel ideas in mobile robot localization and map building. The key contributions are the proposal of a new technique to fuse visual data; the formulation of new algorithms to concurrent localization and map building that take into account temporal landmark quality; new theoretical results on the degree of reconstruction possible when building maps from fully correlated observations; and the necessary techniques to palliate partial observability, partial controllability, and the nonlinear e.ects when solving the simultaneous localization and map building problem.



Author keywords

mobile robotics, localization, state estimation

Scientific reference

J. Andrade-Cetto. Environment learning for indoor mobile robots. UPC, 2003.