Disparity Refinement With Stability-Based Tree For Stereo Matching

In stereo vision,the most challenging topic is stereo matching which extracts disparity information from a pair of stereo images,i.e.a pair of left and right views.Most of the recent algorithms refine the cost value or aggregation step to enhance the performance,while disparity refinement,which plays a significant role in improving disparity quality,attracts far less attention.Disparity refinement can remove all kinds of outliers in the original disparity map and greatly improve the accuracy with low cost.This thesis describes a disparity refinement method with stability-based tree.Existing methods fail to remove large error parts in disparity map since they cannot provide reasonable support regions for error parts.We observe the feature of bad support regions and summarize it as unstableness.By developing stability-based tree to evaluate and reconstruct support regions for error parts,this method achieves effective performance in removing outliers.This approach further improves the quality of raw disparity map in stereo matching,which makes the disparity maps produced by local methods comparable to the global ones.This thesis also provides an optional improvement for extending local methods to realworld road scenes.The difficulties for local methods in road scenes have been discussed.Solutions which contains super pixels as well as HSV color transform have been designed for speeding the process and handling broken texture region,illumination problem,etc.Experiments are done over both the well-known Middlebury data-set and KITTI data-set.The experimental results demonstrate that the presented method reduces more than 70% aggregation time compared with traditional tree method without loss of accuracy in Middlebury dataset and 85% in KITTI dataset.It also outperforms existing disparity refinement methods in removing large error parts.