Abstract: In order to improve the high complexity problem of SIFT, SURF algorithm simplifies and approximates the DoH（Determinant of the Hessian）. This improvement not only guarantees the stability of the algorithm, but also increases the calculation efficiency. But when the SURF simplifies the DoH gaussian second order differential template, some of the image gradient information are lost. Therefore, a Gradual SURF（G-SURF） operator is proposed in the this paper, and the gradient information is added in the process. Experimental results show that the proposed improved SURF operator can get a more stable effect and improve calculation complexity at the same time.

Key words: stereo matching, corner detection, Speeded Up Robust Feature（SURF）, Scale Invariant Feature Transform（SIFT）

摘要： SURF算子为了改善SIFT的计算复杂度高的问题，简化和近似了DoH（Determinant of Hessian），这样不仅保证了算法结果的稳定性，也提高了计算效率。但是SURF这样的近似简化过程，损失了图像中的一些渐变信息。对SURF算子进行了改进，在其处理过程中加入了渐变的信息。实验结果表明，提出的G-SURF（Gradual-SURF）算子可以获得更稳定的效果，并且同时计算复杂度也有所改善。

关键词: 立体匹配, 角点提取, 加速稳健特征（SURF）, 尺度不变特征转换（SIFT）