A major portion of the research efforts of vision community has been given under situation not that of complex environment, noise and illustration but that of ideal. This situation is common in a visual navigation system. In addition, mobile robot navigation needs acquiring positions of obstacles in real time. In order to satisfy these aspects of vision system, a suitable image processing is demanded. In this thesis, indoor images are acquired by binocular vision, which contains various shapes of obstacles. From these stereo image data, in order to detect location of obstacles, it is required to match obstacles between left and right image excluding the environment. It is required to remove noises and to get the precise positions of obstacles.
First, assuming that there is sufficient straight line edge, this thesis presents analysis of conventional stereo techniques using line fitting, and proposes modified Hough transform which results in faster stereo matching. Second, we present that improved correlation matching method enhance the speed of arbitrary obstacle detection. It results in faster, simple matching, robustness to noise, and improvement of precision. Experimental results for both methods under actual surroundings are presented to reveal the performance.