Top View Image Stitching Algorithm Based on 3D Modeling and Region Compensation
Wei Li, Haocheng Zhu, Mingjin Li, Xue Han · 2025
To address the challenges of image deformation and region loss in top-view image stitching scenarios, a novel algorithm based on three-dimensional modeling and region compensation is proposed. This approach aims to effectively fuse images captured from different cameras with varying tilt angles. A new image transformation algorithm is introduced to mitigate lateral bending issues when aligning images from different viewpoints. This algorithm utilizes camera coordinate parameters to construct a three-dimensional model, automatically generating vertex coordinates for point-to-point perspective transformations on the original images, resulting in quadrilateral transformation outputs that enhance the fidelity of the stitching process. Furthermore, to tackle the issue of missing regions in fused images typically resulting from conventional stitching algorithms that preserve only unilateral areas, a region compensation algorithm is developed. This algorithm merges the mask of the missing region on the opposite side of the stitch with the original stitch mask, leading to a complete fused image without any missing sections. Experimental results indicate that the root-mean-square error of the vertical angle of line segments produced by this algorithm is reduced by 89.87% and 90.28% compared to the AutoStitch algorithm and the traditional SIFT+RANSAC algorithm. In addition, the area of pre-served regions and the entropy of image information after fusion are improved by 10.40% and 8.65%, respectively, relative to traditional stitching techniques. This advanced algorithm effectively meets the requirements for monitoring and target detection in construction site applications and similar contexts.