面向大规模SFM的快速鲁棒特征跟踪算法

发布时间：2018-11-20 17:01

【摘要】：运动恢复结构(SFM)是指通过分析二维图像序列恢复三维结构信息的过程,在计算机视觉的多种应用中起着重要的作用。特征跟踪是大规模SFM的核心组成部分,但现有的多视图特征跟踪算法在鲁棒性和效率上还存在不足,为解决这一问题,提出了一种快速和鲁棒的特征跟踪(FRFT)算法。首先,采用AGAST进行特征点检测,并使用图像矩为AGAST特征定义主方向,为构造旋转不变的描述子奠定基础;其次,在差分高斯金字塔空间内,根据中心点与邻域像素之间的差值构造特征描述子,避免光照和尺度变化对特征匹配的影响;再次,为了提高特征匹配效率,对特征集合进行聚类,采用KD-Tree加速特征匹配,提高算法的时间效率;最后,采用4种方式对FRFT算法进行验证,并与现有经典算法进行比较。实验结果表明,FRFT算法在鲁棒性和时间效率方面均优于现有经典算法。
[Abstract]:Motion recovery structure (SFM), which plays an important role in many applications of computer vision, refers to the process of restoring 3D structure information by analyzing 2D image sequences. Feature tracking is the core part of large scale SFM, but the existing multi-view feature tracking algorithms are still insufficient in robustness and efficiency. In order to solve this problem, a fast and robust feature tracking (FRFT) algorithm is proposed. Firstly, AGAST is used to detect the feature points, and the image moment is used to define the main direction for the AGAST feature, which lays the foundation for constructing the rotation-invariant descriptor. Secondly, in the differential Gao Si pyramid space, the feature descriptor is constructed according to the difference between the center point and the neighboring pixel to avoid the influence of the illumination and the scale change on the feature matching. Thirdly, in order to improve the efficiency of feature matching, the feature set is clustered, and KD-Tree is used to accelerate the feature matching to improve the time efficiency of the algorithm. Finally, four methods are used to verify the FRFT algorithm, and compared with the existing classical algorithms. Experimental results show that the FRFT algorithm is superior to the classical algorithms in robustness and time efficiency.
【作者单位】：合肥工业大学计算机与信息学院;
【基金】：国家自然科学基金(61370167) 安徽省自然科学基金(JZ2015AKZR0664)项目资助
【分类号】：TP391.41

【参考文献】