ObjectiveWhether image stitching algorithm is effective or not depends on image registration and image blending. In the stitching of forestry aerial images, misalignment caused by registration error and ghost caused by disparity could be diminished or eliminated. Optimal seamless-line algorithm's general effect is better than other blending algorithm. But it doesn't involve orthographic effect into stitching result. When UAV flies in a very low altitude, the ratio of tree height to flight height is larger, which impedes to obtain orthographic panorama. To get orthophoto of image stitching, traditional methods always use DEM data to compute digital differential rectification. However, DEM data and plant height are variable and difficult to measure. Besides, the recovered terrain as well as plant height have error with ground truth, which leads DOM to generate distortion and obscure in some local areas. Therefore, we proposed a novel improved optimal seamless-line blending algorithm using image information instead of DEM data, which can reserve orthographic projection in the final panorama.

MethodFirstly, SURF was extracted, matched and filtered. RANSAC algorithm was used to compute homographic transform matrix in order to obtain the overlap area. Secondly, we constructed energy function, including distance difference between each pixel in overlap area and its image center, based on the ideology that distance difference can reflect the orthographic effective. Each item's dynamic parameter was set to weigh color, structure and distance. Dynamic programming was applied to seek optimal seamless-line. In the end, multi-band blending was used to generate seamless panorama.

ResultThe image datasets used in experiment varied from land region and UAV altitude. The stitching result of two neighbor images and airline image series showed that our method can reserve orthographic projection whose visual effect was better than current optimal seamless-line algorithm. In the town panorama test, our method was better than DOM generated by Pix4D in local regions.

ConclusionWe proposed a novel improved optimal seamless-line algorithm using image information instead of DEM data. The algorithm can diminish misalignment and ghost, and can reserve orthographic projection in the final panorama, whose visual effect is similar to DOM. Besides, the method does better in retaining sharp edge of object than DOM generated by commercial software. The method helps to compute forest canopy density as well as vegetation coverage area more accurately, thus contributes to tracking and recognizing animals and plants more precisely. In addition to forestry, the algorithm can be extended to other types of aerial images, such as town images, and orthographic projection should be reserved in the panorama.