Abstract: Stereo matching is the most important step in binocular vision, and it is also the most difficult problem to solve. The traditional stereo matching to get the regional disparity map is just passable, for the characteristics of the non-texture regions are difficult to retain the depth map discontinuing, so getting the 3D information of the target is slow and inaccurate, Therefore, a new SIFT image matching is proposed, firstly, making the epipolar constraint on the left and right image; secondly, selecting the ROI of target from the left of binocular images, and reducing running time by reducing the dimension of feature vectors and accelerating the matching speed by using BBF algorithm based on KD tree, and finally, removing the false matching by using RANSAC algorithm. The experiment results show that the improved SIFT algorithm can get the target's feature points quickly and accurately, so the 3D coordinates can be calculated by the triangulation method speedy.