Research On Multi View Dense Point Cloud Reconstruction Based On Low Altitude Images

With the rapid development of sensor technology and spatial information technology, people's applications and needs about 3D space information became higher. Recently years, the research of ulitizing Laser Scanning System to achieve the 3D data about object and scene has aroused. As a new sensor, the Laser Scanning System is too expansive to widely popularize and applicate in short times. The reconstruction of 3D model through stereo image matching has lower cost, and it can keep the spectrum and texture information of object and scene, which is deficient by Laser Scanning point cloud. Stereo matching using low altitude remote sensing images as source data is one of the main ways to reconstruct large-scale scene.Compared with the traditional remote sensing platform, the low-altitude remote sensing platform has lots of superiorities like low altitude and low cost of flight, fast response speed, high resolution image and can make up the deficiency which the traditional remote sensing plateform is difficult to shoot real-time images due to the restriction of the weather conditions. However, the low altitude remote sensing platform is easy to be affected with airflow since the plateform has small size and low altitude flight, and the flight attitude of the plateform is not stable, which has brought new challenges to image stereo matching.Aiming at the characteristics of the low altitude images, the researches of this dissertation focuse on multi view dense point cloud reconstruction through low altitude images. The major innovations of this dissertation include the following aspects:(1) A least squares image matching method based on non-fixed initial patch. The method is based on LSM algorithm. By using the geometry relationship of the point cloud in object space coordinate system to calculate approximate normal vector of local tangent plane, and by ulitizing this approximate normal vector to establish the initial patch. The propose of the method is to optimize stereo matching by building up least squares error equations with the pixels which the non fixed initial patch project on the images and calculating the optimal solution with least squares adjustment. The experimental results has illustrated that this method is superior to traditional patch based LSM method both in the efficiency and the accuracy of image matching.(2) A vertical patch based least squares image matching method. Aiming at the problem about image matching on object discontinuities, the method sets up a pair of mutually perpendicular patches, and restricts the least squares image matching by the relationship between the vertical patches and the images, which can address the error matching that the local tangent patches on object discontinuities are not exist. The superiority of the method in matching accuracy is proved by experiments.(3) A multi-view stereo matching method based on low altitude image is presents to focus on the three challenges in reconstruction 3D model by stereo matching using low altitude remote sensing images at matching accuracy, point cloud density and the model integrity. The method ulitizes PMVS result as seed patches and takes advantage of projection relationship between the pixels in image windows and the object points on patches to segment and expand the seed patches, then an improved patch based least squares image matching method is employed to optimize the point cloud point by point. The method takes advantage of redundant measurements of multi-image to weaken the influence about noise and occlusion. The experiments has illustrated that the method has kept the advantages of multi-view stereo matching on matching accuracy and model integrity, at the same time, the density of point cloud is greatly improved. The experiment has also proved that the reconstruction technology based on low altitude remote sensing images has a denser point cloud data than Laser Scanning System when the ground resolution of the images is high enough.The methods in this dissertation are applied to the urban 3D reconstruction.