To solve problems of the difficulty to meet the real-time requirements and the low matching accuracy of existing local stereo matching algorithms at some special regions, such as weak textured surfaces and the discontinuity boundary of depth, a dense stereo matching algorithm based on cross-scale guided image filtering is proposed. An image segmentation technology is used to realize pre-segmentation of stereo images and the aggregation radius of pixels in the segmented region is obtained. This radius is used as a guide, and kernels with three different sizes are used to carry out filtering in the cost space of stereo image. The regularization term is introduced to ensure the consistency of the aggregated cost, so as to obtain a more efficient aggregate cost. A simple and efficient winner-take-all strategy is used to obtain the initial disparity. The experimental results based on Middlebury test bench show that the proposed algorithm has both real time capability and high efficiency.