车载立方体全景影像匹配点的粗差检测方法

A Gross Error Detection Method of Vehicle-borne Cubic Panoramic Image Sequence

针对低匹配内点率情况,利用随机抽样一致性算法(RANSAC)估计车载全景序列影像的极几何模型不稳定,造成大量匹配点的粗差无法检测或误检测问题,提出了一种基于多约束条件的粗差检测方法。以冗余粗差为约束条件,采用SIFT和最近邻匹配方法获取独立随机匹配点,并构建特征光流矢量。利用光流幅度和方向直方图统计结果,融合极线、尺度和天空点约束条件实现全景影像匹配点的粗差检测。通过不同数据的实验分析,在短基线条件下,可以有效地检测出大部分由纹理重复性、尺度变化和运动物体产生的匹配粗差点。与传统方法比较,本文方法可获得更高的匹配正确点数和正确率,尤其在场景复杂造成的低内点率情况下,算法表现较为稳定。

Because the epipolar geometry model estimation of panoramic images is unstable under the low match inlier ratio cases, large numbers of outliers or errors cannot be detected using RANSAC method. A new gross error detection method based on multiple constraints is presented for vehicleborne panoramic image sequences. First, the initial matching points are extracted using SIFT and nearest neighbor matching, then independent random matching points are constructed by redundant gross error constraints. Second, the movement relationships between panoramic images are approximately expressed by the histogram statistics of optical flow magnitude and direction, which can effectively improve the matching inlier ratio. Finally, the epipolar geometric constraint, scale constraint and sky point constraint are used for gross error detection. Several panoramic images were selected and used for experiments. An analysis and comparison were carried out on these data. The results show that the proposed method works well in short-baseline conditions for the number and accuracy of correct matching points, especially for complex scenes in low inlier ratio cases. The algorithm performance is relatively stable, and provides better constraint for gross errors usually caused by repeated textures, scale changes, and moving objects.