关键词: 尺度不变特征转换算法, 立体视觉, 特征点匹配, 全局信息, 特征描述符

Abstract: The high complexity and long computing time of Scale-Invariant Feature Transform(SIFT) algorithm cannot meet the real-time requirements of stereo matching. And the mismatching rate is high when an image has many similar regions. To solve the problems, an improved stereo matching algorithm was proposed. The proposed algorithm was improved in two aspects. Firstly, because the circular has natural rotation invariance, the feature point was acted as the center and the rectangle region of the original algorithm was replaced by two approximate-size concentric circle regions in the improved algorithm. Meanwhile, the gradient accumulated values of 12 directions were calculated within the areas of the inner circle and the outer circle ring respectively, and the dimension of the local feature descriptor was reduced from 128 to 24. Then, a 12-dimensional global vector was added, so that the generated feature descriptor contained the SIFT vector based on local information and the global vector based on global information, which improved the resolving power of the algorithm when the images had similar areas. The simulation results show that, compared with the original algorithm, the real-time performance of the proposed algorithm was improved by 59.5% and the mismatching rate was decreased by 9 percentage points when the image had many similar regions. The proposed algorithm is suitable for in the case of high real-time image processing.

Key words: Scale-Invariant Feature Transform(SIFT) algorithm, stereo vision, feature point matching, global information, feature descriptor