Remotely controlled mobile robots have been a subject of interest for many years. They have a wide range of applications i n science and in industries such as aerospace, marine, forestry, construction and mining. A key requirement of such control is the full and precise knowledge of the location and motion of the mobile robot at each moment of time.

This thesis presents a vision-based location tracking system suitable for autonomous vehicle navigation and guidance in unknown environments. The system includes a trinocular vision head that can be mounted anywhere on a navigating robot. Consecutive sets of triplet 2D images are used to determine the 3D location and the 3D motion parameters of the robot at each frame. By selecting only the most informative points of each image, using a feature detection algorithm, faster performance is achieved. The use of 3 cameras improves the accuracy and the robustness of the system. By tracking the 3D location of world features within a multi-stage tracking approach, the location of the observer camera is estimated and tracked over time. The motion with 6 DoF is found via a least squares minimization method. A Kalman filter implementation is used to optimize the 3D representation of scene features in order to improve the accuracy of the overall system.

The system introduces several novel contributions for vision-based trajectory tracking. Firstly, it presents a new binary corner detector that can automatically detect the most informative and reliable scene feature points. This feature detector performs 60% faster than the most common method, the Harris corner detector, used in vision-based tracking applications. Secondly, it compensates for an inherent source of inaccuracy in the camera geometry. By identifying and matching features in raw images, and then unwarping matched data, accurate 3D world feature reconstruction is achieved. Thirdly, through a two-stage search and tracking design, similar 3D world feature points axe identified and tracked over time. This design, by itself, improves the accuracy of the estimated motions at each time by up to 10%. Finally, it improves the accuracy of the 3D trajectory tracking system with 6 degrees of freedom in natural environments.

Results of the application of our method to the trajectory tracking problem in natural and unknown environments are reported to validate our approach, hypotheses and methods. The cumulative translational and rotational errors of the system are less than 1% for the studied examples.