Research On Multi-view Oblique Aerial Image Matching

Compared to conventional single-lens photogrammetry, multi-view oblique aerial photogrammetry which can obtain the lateral textures of tall buildings contributes to the entire and refined 3D modeling and draws more and more attention. Image matching is the core stage of digital photogrammetry, whose quality has direct influence on the following relative orientation, aerial triangulation and DSM generation. This paper is mainly focused on imaging characteristics of multi-view photogrammetry, matching constraint, feature-based matching and dense matching. The main work and innovation are as follows:1. As the theoretical foundation, the data structure of oblique images and imaging geometry of oblique photogrammetry are introduced. Hierarchical image pyramid matching strategy, homography constraint and epipolar geometry constraint are adopted to restrain oblique image matching. The constraint effect of homography and epipolar geometry are compared.2. An equivalent nadir- rectification methodology is proposed to make compensation for geometric distortion. Experimental results show that there is only scale difference between oblique images after the equivalent nadir- rectification and the relative nadir images. The SIFT matching effect on equivalent nadir- images is better than ASIFT effect on the original oblique images, which certifies that the equivalent nadir- rectification method can compensate for geometric distortion to some extent.3. The absolute tilt measure and relative tilt measure calculation methods are designed to measure the tilt degree of multiple-view images quantitatively. The relationship between ASIFT matching effect and image tilt measure is studied. The experiments show the relative tilt measure calculated from this method takes on reverse proportion to ASIFT matching ratio, which certifies the calculation method of the relative tilt degree.4. An adaptive hierarchical dense matching method between nadir image and oblique image is put forward. As for matching two images of different views, the local rectified compensation for geometric distortion based on geometric transformation is implemented before area-based matching. Homography constraint, epipolar geometry constraint and consistent disparity constraint based on irregular triangular network are applied. This algorithm is entirely based on image space without the utility of exterior orientation elements and takes adequate consideration of the time-consuming problem of resampling and coordination projection. Experimental results indicate the algorithm can effectively compensate for geometric distortion of different-view images and achieve reliable per-pixel dense matching.5. Different matching algorithms are designed to match continuous image sequences of different perspectives. As for matching forward-view image sequences, the equivalent nadir rectification is made first and local scale normalization is implemented following this stage. The last matching step can be achieved by area-based matching. For the continuous nadir-view, left-view and right-view image matching, the dominant image distortion is resulted from perspective rotation, which can be compensated by local scale normalization and quasi-epipolar rectification based on reliable matching features. The final step is area-based matching or semi-global matching on quasi-epipolar images. The experiments judge the quasi-epipolar rectification effect through calculating the epipolar constraint closure error and the vertical parallax on the quasi-epipolar images. And the matching results and point clouds show this algorithm can succeed in matching continual neighboring images.