Modelling the environment of an exploring vehicle by means of stereo vision

This dissertation describes research involving vision techniques which would be useful in an autonomous exploring vehicle, such as a Mars rover. These techniques produce a description of the surroundings of the vehicle in terms of the position, size, and approximate shape of objects, and can match such scene descriptions with others previously produced. The information produced is thus useful both for navigation and obstacle avoidance. The techniques operate by using three-dimensional data which they can produce by means of stereo vision from stereo picture pairs or which can be obtained from a laser rangefinder. The research thus divides conveniently into two portions: stereo mapping and three-dimensional modelling and matching. The stereo mapping techniques are designed to be suitable for the kind of pictures that a Mars rover might obtain and to produce the kind of data that the modelling techniques need. These stereo techniques are based upon area correlation and produce a depth map of the scene. Emphasis is placed upon extraction of useful data from noisy pictures and upon the estimation of the accuracy of the data produced. Included are the following: a self-calibration method for computing the stereo camera model (the relative position and orientation of the two camera positions); a high-resolution stereo correlator for producing accurate matches with accuracy and confidence estimates, which includes the ability to compensate for brightness and contrast changes between the pictures; a search technique for using the correlator to produce a dense sampling of matched points for a pair of pictures; and the computation of the distances to the matched points, including the propagation of the accuracy estimates. The three-dimensional modelling and matching techniques are designed to be tolerant of the errors that stereo mapping techniques often produce. First, a ground surface finder tries to find a set of points that form a well-defined smooth surface that lies below most of the other points. Then, by using this knowledge of the ground surface and knowledge of the camera viewpoint that produced the points in the scene, an object finder approximates the objects that are above the ground by ellipsoids. Finally, a scene matcher can use the descriptions of scenes in terms of ellipsoidal objects. By using a search pruned by using probabilities obtained by means of Bayes' theorem, it determines the probability that two scene descriptions refer to the same scene and the linear transformation needed to bring the two scenes into alignment. These techniques have been tried on stereo pictures of the Martian surface taken by the Viking Lander 1. The object finder was able to locate rocks fairly successfully, and the scene matcher was able to match successfully the resulting scene descriptions. Examples of these results are shown.