FAST AND PRECISE BINARY IMAGE DESCRIPTOR FOR AUTONOMOUS VEHICLE VISUAL LOCALIZATION

Abstract

Autonomous vehicle self-localization by scene matching under extreme environmental changes has been among the most challenging problems in robotics and computer vision in the last few years. Large dynamic illumination changes during day hours and appearance changes between year seasons are the major difficulties of this problem. This thesis presents: 1) a new binary image descriptor addressed as ”Extended Local Difference Binary” (ELDB), which is an extension to the state-of the-art Local Difference Binary (LDB) image descriptor, and 2) a new algorithm for vehicle visual localization under extreme environmental changes that uses Multi-Hypothesis Markov Localization (MHML) as a data fusion algorithm, and uses ELDB for image matching. Experimental results presented in the thesis show that ELDB has better image matching accuracy and computational efficiency than LDB, and that the proposed vehicle visual localization algorithm is faster and more accurate than other state-of-the-art algorithms.