Based on the existing graph cut algorithms, a stereo matching method which is based on adaptive watershed algorithm and non-parametric depth smoothing model to create an energy equation of graph cut is designed. A novel adaptive local threshold method is proposed, which is applied to region integration by combining the watershed with Prim′s algorithm. The proposed method selects the pixels with same brightness as a feature vector and forms pixels group, thus two or more images can be matched in the pixels group layer of feature region, then the amount of data can be greatly reduced. When the energy equation is minimized, pixels groups layer can be used to optimize the α-expansion algorithm, and reduce the running time. Experimental results indicate that the error match ratios of all regions, non-occluded area and depth discontinuity regions are all less than 8.5%. The proposed algorithm is in the place of 19 among all the 135 algorithms in Middlebury testing platform.