基于特征点的立体匹配算法研究

【作者】 李鹏；

【导师】 罗桂娥；

【作者基本信息】 中南大学 ， 电子科学与技术， 2012， 硕士

【摘要】 立体匹配就是对多幅图像中的对应点进行匹配,它是获得图像的深度信息最关键的步骤。立体匹配是计算机视觉领域中的一个重要研究方向,同时也是立体视觉中的一个瓶颈问题。研究一种鲁棒性好、精确度高、性能稳定和适用性强的立体匹配方法具有重要的理论意义和实用价值。本文作者对立体匹配中特征点的提取算法和特征描述子的生成方法做了较深入的理论研究和大量的实践工作。本文对Harris特征点提取算法进行改进,提出了一种精确度和实时性更优越的特征点提取算法,此算法首先对图像所有像素点进行一次简单的初始选择,排除一部分普通像素点,得到数量相对较少的候选角点集,后续的Harris角点提取范围从整幅图像缩小为该候选角点集,从而降低了算法的时间复杂度。利用圆的旋转不变性,本文还对SIFT特征描述子的生成方法进行了改进,该方法首先构造以特征点为中心的五个同心圆环,通过同一圆环内的所有像素点的八个方向梯度累加值排序生成的特征向量来表征特征点信息,这样生成的描述子本身就具备旋转不变性,特征向量维数的降低为后续的匹配节省了时间。最后将改进的算法分别引入到立体匹配中的特征提取和特征描述过程中,提出了一种基于Harris角点和SIFT特征描述子的立体匹配算法。该算法首先采用改进的Harris角点提取算法对预处理后的图像对进行角点提取,然后用改进的方法生成特征描述子,再利用描述子间的相似性度量进行匹配并依据约束准则删除误匹配,得到匹配结果。实验结果证明该算法在实时性和准确性方面的性能都有较大提高。更多还原

【Abstract】 Stereo matching is to match the corresponding points in pieces of different images, and it’s the most critical step of obtaining the depth information of the images. It’s an important research direction in computer vision field. At the same time, it’s bottle-neck problem in the research of stereo vision. It’s a great challenge to investigate a robust, precise, stable and applicable stereo matching algorithm, which is both theoretically and practically valuable.This paper finished deep theoretic research and sufficient practical work on feature point extraction algorithm and the method how to structure feature descriptor for stereo matching. This paper improved the Harris feature point extraction algorithm, which presented a new feature point extraction algorithm. In order to obtain a relatively small amount of corner points, this algorithm makes an initial selection for all pixels of the images, and removes a part of ordinary pixel firstly. In this way, the range of extraction of Harris corn points is reduced, so the time complexity of the algorithm is reduced, which improves accuracy and real time of the algorithm. By using the rotational invariance of the circle, generation method of the SIFT feature descriptor is improved. Firstly, this method structures five circle areas, and each center is feature point. Then, characterization of feature points information with eigenvectors is generated by accumulate the gradients of eight directions for all Pixels in the same circle. The descriptor generated by this way has rotational invariance intrinsically. In addition, it is useful for the consecutive matching to reduce the dimension of the eigenvectors. The last, a stereo matching algorithm that is embedded in Improved algorithm is base on Harris comer and SIFT feature descriptor. Firstly, the algorithm extracts the corner points with improved Harris corner detection operator, then it generates the feature descriptors, at last, it matches to feature points by using the similarities of descriptors, and false matches are removed according to the constraint criteria. The experimental results are proved that the algorithm presented in this paper can improve the accuracy and real time.更多还原