The visual odometry method is adopted based on Scale Invariant Feature Transform(SIFT) to get stereo corresponding points for motion estimation between every two binocular frames. Then the rotation and translation matrix is employed to invert the moving path of the robot. SiftGPU is applied to accelerate computing so that real-time visual odometry system can be achieved. RANSAC algorithm is adopted to eliminate mismatching interference. Experimental results show that SIFT algorithm with strong scale and affine transform invariance can get more accurate path inversion results; and GPU graphics acceleration will achieve real-time visual positioning.