基于改进Patchmatch及切片采样粒子置信度传播的立体匹配算法

Stereo Matching Algorithm Based on Improved Patchmatch and Slice Sampling Particle Belief Propagation

针对立体匹配时视差不连续区、倾斜平面及非前向平行平面误匹配较高的问题,提出了一种基于改进Patchmatch及切片采样粒子置信度的立体匹配算法.定义了具有边缘特性的Patchmatch相似性函数,并建立基于Patchmatch的非前向平行平面视差平面估计模型.利用粒子置信度传播代替原有的最近邻搜索,使用较少的粒子近似目标分布,并采用切片采样马尔可夫链蒙特卡罗方法解决传播过程中粒子重采样更新问题.Middlebury图像数据集测试表明,该算法能够降低视差不连续区域的误匹配,有效地提高了倾斜平面及非前向平行平面图像的匹配精度.

The high erroneous results of the stereo matching occur at the following three cases where there are depth discontinuity region, the slanted surface or the non-fronto-parallel surface. A stereo matching algorithm was proposed based on the improved Patchmatch and slice sampling particle belief propagation. An edge-preserving similarity function of the Patchmatch was defined. Then, a model of the depth estimation for the non-fronto-parallel surface was introduced. The nearest neighbor search was replaced with the particle belief propagation, and the target distribution was approximated with a finite set of particles. At the same time, the sampled particles from the belief distribution was typically done by using slice sampling Markov chain Monte Carlo method to solve the particle update problem. The experiments on the Middlebury indicate that the mismatching at the depth discontinuity region can be reduced, and the match accuracy for the slanted surface and the non-fronto-parallel surface can be improved.