An Efficient Hierarchical Stereo Matching Approach Combining Daisy-LBP Descriptor with Regional Plane Constraints

Although significant advancements have been made in stereo vision over the past decade, reliable depth estimation remains a challenge, especially in regions with repetitive or poor textures and at depth discontinuities. To address these challenges, an efficient hierarchical stereo matching approach is introduced, leveraging the strengths of the PatchMatch filter (PMF) and a multi-resolution pyramid strategy. Firstly, a stereo image pyramid is constructed using downsampling parameters, and each layer is segmented by the Simple Linear Iterative Clustering (SLIC) algorithm to build a superpixel-level adjacency graph. Secondly, a robust cost function is designed by integrating the Daisy-LBP descriptor and image gradients, based on which a hierarchical multi-resolution cost volume is constructed within the pyramid framework. Thirdly, a hierarchical stereo matching strategy is embedded into a robust PMF algorithm. The strategy begins processing the lowest-resolution image pair, applying cross-based local multipoint filtering (CLMF) for cost aggregation and refining pixel-wise disparity planes using regional plane constraints. At higher-resolution layers, disparity planes are initialized using the optimal results from the preceding layer. This iterative process continues until it reaches the original-resolution image pair. Lastly, the robust PMF algorithm integrates multi-resolution cost aggregation with prior disparity planes at the original resolution to further improve robustness in poor-texture and repetitive regions. Experimental evaluations conducted with the Middlebury and KITTI benchmarks demonstrate that the proposed approach enhances both matching accuracy and robustness, particularly in challenging regions and at depth discontinuities, outperforming comparative methods.