| The ability for a robot to make intelligent autonomous decisions within a previously unknown environment is directly dependent on its ability to both build a usable representation of its surroundings (a map) as well as localize itself with respect to that representation. In stereo video scene reconstruction the typical pipeline for addressing this task is as follows: capture a stereo image pair, determine correspondences between points in the images to generate a set of 3D points, estimate the motion of the camera by registering a subset of the matched 3D points across time and then combine the 3D data into a consistent map. This thesis explores guiding the search space for correspondence matches based on the state information present within the stereo video reconstruction system. Prior 3D surface models are integrated with a camera motion model assuming a static environment. This provides a prior for the expected disparity surface which is used to inform the subsequent stereo matching process, enhancing the stereo matching task both in terms of accuracy and efficiency. Experimental validation with simulated, real world ground truth datasets, as well as underwater and terrestrial stereo image sequences are used to demonstrate the benefits of this approach. |
