Passive depth estimation using stereo vision, an experimental study

Abstract

This master s thesis is an experimental study on passive stereo techniques for retrieving 3D scene information. The stereo camera system is tested for finding the position of a platform relative to the stereo camera where the measurements are used for navigating an autonomous docking system installed on offshore vessels.

The accuracy of passive stereo techniques depends on having precise knowledge of the cameras position, orientation and internal parameters. Calibrating these parameters is a key part of the success of the stereo camera. For a point in the scene imaged from two different known view points, the displacement of the point reprojection in the images are inversely proportional to depth and can b e used to compute the 3D coordinate. The problem of establishing correspondences in the image pair is a difficult task and there exist several algorithms for solving the correspondence problem. In this thesis, we have evaluated the accuracy of the stereo camera system for computing 3D information of the scene and proposed an optimal baseline length to improve the accuracy. Six different of algorithms have been tested for solving the correspondence problem for this application. The algorithms considered are the correlation based methods: Sum of absolute differences, squared sum of differences and normalized cross correlation. Together with the feature based methods: Harris corner detector, scale invariant feature transform and speeded up robust features. The performance of each algorithm is compared through a field experiment to conclude on the b est one. Based on the results of the experiment, the correlation based methods proves to be best suited for this application where the sum of squared differences is considered as the superior algorithm for solving the correspondence problem.