Binocular stereo matching algorithm is real-time and accurate in the binocular vision servo control system, which is very important for the accurate positioning and control of the robot. Based on the study of some advanced stereo matching algorithms, this paper introduces an algorithm of local stereo matching - ELAS (Efficient Large-scale Stereo) algorithm. This local stereo matching is based on processing of the parallax plane. The algorithm overcomes the shortcoming of these current mainstream algorithms that only by setting the maximum disparity value range, can we obtain the best disparity maps. In consideration of processing time and processing effect, this algorithm has best performance. The principle of ELAS algorithm is analyzed thoroughly in this paper. The defects are also put out. Aiming at the problem that the ELAS algorithm does not work well, the improved ELAS algorithm is studied according to the principle of disparity continuity. The disparity map is processed using the guided filter. Experiments show that the improved ELAS algorithm improves the performance of the original algorithm significantly. Without considering the real-time performance, combined with the above-mentioned filters, the processing effect can be close to the real disparity map, which is of great significance in 3D reconstruction and visual servoing.