A Reliable Multi-View Image Matching Approach Based On The Moving Z-Plane Constraint

Reliable image matching is an essential and difficult task in digital photogrammetry and computer vision. The traditional image matching is restricted to the imaging abilities of stereo sensors, which based on the matching of“single stereo-pair”, and therefore is a challenging and“ill-posed”problem. Along with the increasing use of new digital sensors, it becomes more and more easier to acquire large overlap digital images covering the same area, the multi-view image matching approach has attracted wide interests in both photogrammetry and computer vision.Multi-view image matching has three significant advantages: Firstly, it improves the reliability of matching utilizing the comprehensive multi-view image information; Secondly, it effectively solves the relative-occluded problem of images; and thirdly, it can selectively match the multiple images on-demand. However, up to now the existing multi-view image matching methods only employ the first advantage, there is very few research about another two characteristics, and the matching strategy is also relative inflexible. Focusing on the serious occlusion problem in city images, this dissertation proposes a reliable multi-view image matching method based on the moving Z-Plane constraint. Following a constraint plane of grid partitioning in object space, this method adopts the image interest points and the“groundel”(ground element) in object space as the matching primitives for feature matching and dense matching. Simultaneously, it carries on the selective matching for multi-view images depending on the projection rays of matching primitives to avoid the occlusion effect in the matching. This method suffices the conception of true multi-view image matching, which can deal with any number of images, and any degree overlap areas in to-be-matched images.The main contents of this dissertation are as follows:⑴Image pre-processing and extraction of interest points. In image pre-processing, this dissertation adopts the adaptive smoothing filter to reduce the image noise and the Wallis filter to enhance the image features. It also compares several extraction operators of interest points (Moravec operator、Forstner operator、Harris operator) on location accuracy and numbers of interest points through experimental analysis, and finially employs the Forstner operator to extract interest points in overall images.⑵Interest point matching. A new constraint mode in object space for the simultaneously multi-view image matching is introduced. By moving the matching plane to different elevation positions so called Z-Plane, it constraints the range of projection rays from different images based on the positions of grid cells in the Z-Plane , and then a related hierarchical matching according to the number of viewing rays in the grid cells is presented. Simultaneously, it carries on a selective matching depending on the viewing rays in the grid cells, in order to avoid the matching vulnerability caused by occlusions or interest point not be extracted. In the process of matching, it also assigns the evaluation values to successfully matched grid cells.⑶Dense matching. It takes the valued grid plane after interest point matching as the initial DSM, and as the carrier of regular distributing“groundel”in the object space. Introduces the occlusion detection method into the process of“groundel”matching in the object space. Along with the height-based occlusion detection method and the simultaneously multi-image matching based on vertical line constraint in the object space, it carries on the occlusion detection before the matching of each groundel, and selects the images without occlusions to do dense matching according to the result of occlusion detection, in order to avoid the effect of occlusions and to obtain good matching results in relative-occluded areas, even the occlusion map in every image.⑷A multi-view image matching strategy based on the unfixed reference image is proposed. Focusing on the problem of reference image selection in the process of matching based on gray area correlation, a new algorithm for reference image selection is designed, which selects the nearest image from the ground nadir point as reference image according to the principle of minimum image distortion in nadir point. The reference image is unfixed in the process of matching. The experimental results validate which selects the nearest image from the ground nadir point as reference image under the same conditions can improve the maximum of cross correlation coefficient in gray window with different images.Based on the multi-view image matching method proposed in this dissertation, it carries on the matching experiment utilizing four large overlap UCD digital airbrone images in the same strip, and compares with the result derived from the VirtuoZo System. The experiment results validate the correctness and effectiveness of this method proposed in this dissertation. Through the selective matching toward multi-view images, this method effectively solves the effect of building occlusion, and provides a new effective solution for reliable multi-view image matching.