ACCURATE AND FAST STEREO VISION

Abstract

Stereo vision from short-baseline image pairs is one of the most active research fields in computer vision. The estimation of dense disparity maps from stereo image pairs is still a challenging task and there is further space for improving accuracy, minimizing the computational cost and handling more efficiently outliers, low-textured areas, repeated textures, disparity discontinuities and light variations. This PhD thesis presents two novel methodologies relating to stereo vision from short-baseline image pairs: I. The first methodology combines three different cost metrics, defined using colour, the CENSUS transform and SIFT (Scale Invariant Feature Transform) coefficients. The selected cost metrics are aggregated based on an adaptive weights approach, in order to calculate their corresponding cost volumes. The resulting cost volumes are merged into a combined one, following a novel two-phase strategy, which is further refined by exploiting semi-global optimization. A mean-shift segmentation-driven approach is exploited to deal with outliers in the disparity maps. Additionally, low-textured areas are handled using disparity histogram analysis, which allows for reliable disparity plane fitting on these areas. II. The second methodology relies on content-based guided image filtering and weighted semi-global optimization. Initially, the approach uses a pixel-based cost term that combines gradient, Gabor-Feature and colour information. The pixel-based matching costs are filtered by applying guided image filtering, which relies on support windows of two different sizes. In this way, two filtered costs are estimated for each pixel. Among the two filtered costs, the one that will be finally assigned to each pixel, depends on the local image content around this pixel. The filtered cost volume is further refined by exploiting weighted semi-global optimization, which improves the disparity accuracy. The handling of the occluded areas is enhanced by incorporating a straightforward and time efficient scheme.

The evaluation results show that both methodologies are very accurate, since they handle efficiently low-textured/occluded areas and disparity discontinuities. Additionally, the second approach has very low computational complexity. Except for the aforementioned two methodologies that use as input short-baseline image pairs, this PhD thesis presents a novel methodology for generating 3D point clouds of good accuracy from wide-baseline stereo pairs.