Abstract: In order to improve the real-time performance and robustness of SIFT，this paper proposes a matching algorithm of binocular vision based on multi-scale feature extraction.This algorithm devises two improved ways.One way is presented that the size of template can automatically adjust along with the scale factor changing.The other is that a bidirectional feature points matching algorithm is proposed.By it，the traditional algorithm is optimized from feature point extracting time and matching accuracy.Finally，experiment shows that by the use of the improved algorithm，left image and right one based on binocular vision can be matched quickly and accurately.