An obstacle detection method is proposed based on binocular stereo vision for manipulator. Firstly, images are preprocessed after camera calibration such as graying and distortion correction. Then, aiming at the defect that feature points are extracted mixed with many edge response points by traditional SIFT (Scale Invariant Feature Transform) algorithm, we propose wavelet transform to extract the edge points, and compare the edge points with feature points to further eliminate unstable feature points. Finally, the effectiveness of the algorithm is verified by obstacle detection and obstacle avoiding experiments. The results demonstrate that the improved SIFT algorithm can remove the edge responses that the DOG operator generates effectively, and improve the stability and anti-noise ability of SIFT algorithm compared with traditional SIFT algorithm.