| Geospatial digitization and informatization have helped people in their daily life since Gore’s speech in 1998. Geospatial products such as Google Earth and Baidu Map have brought great benefits to people’s life, especially in navigation and tour. When the digital earth is being developed to the intelligence earth, Chinese government is developing intelligence China and digital city greatly at the same time. Digital Surface Model (DSM) has been paid a lot of attentations, which is one of the most important geospatial products. The DSM generation needs 3D point clouds. The main obtaining ways for 3D point clouds can be classified into image dense matching technology and laser scanning technology currently. Image dense matching technology is able to recover the 3D geospatial information from 2D images according to the theory of intersection of corresponding light ray. The LiDAR system consists of GPS, IMU and laser scanner. It can acquire 3D point clouds actively. Both the dense matching technology and the LiDAR system have advantages and disadvantages respectively. Image dense matching has advantages of dense point density, obvious gray features of edges, low cost of data acquisition, and so on. However, the results of dense mathing are influenced by radiation quality of images, terrain texture and occlusions. The LiDAR system is able to acquire and update 3D point clouds timely, but it is expensive and the density of the point clouds is low. According to the characteristics of dense matching and the LiDAR system, it is able to achieve denser and more accurate 3D points if the two technologies are combined together. So far, there have been various dense matching methods based on control points. However, these methods are still not mature:1. the dense matching methods didn’t take full advantages of obvious gray feature of edges; 2. the methods didn’t consider the large amount of LiDAR outliers on the local sacle; 3. the methods didn’t consider the fusion of LiDAR point clouds and dense matching point clouds. Thus, it needs to research the method of multi-view image dense matcing and point fusion based on LiDAR point cloud constraints. Research about how to make full use of edge features during matching, how to filter outliers in LiDAR point clouds, and the fusion of dense matching points and LiDAR points has great values in both theory and practice.This paper aims at researching the method of multi-view image dense matcing and point fusion based on LiDAR point cloud constraints. The main work is as follows:1) A method for choosing optimal stereo pairs based on multi-constraints is proposed. Stereo pairs are basic models for dense matching. Choosing optimal stereo pairs can not only avoid redundant calculations, but also guarantee the matching accuracy. Firstly, this paper compares the lengths of base lines between the base image and the other image, and then confirms the potential stereo pair sets; secondly, calculate the intersection angle of image normal vectors between the base images and the image in the potential stereo pair set, and eliminate the image with a large intersection angle; Finally, acquire feature matching points from the base image and the images in the potential stereo pair set, and decide the optimal stereo pair set according to the number of matching points and the intersection angles.2) A new image-guided non-local stereo matching method with three steps optimazition is proposed. Traditional dense matching methods can be divided into local methods and global methods according to the cost aggregation way. The local methods suppose the pixels with similar intensities should have the same disparity. The cost aggregation is guided by image intensities. The global methods define a global energy function. The disparity image is the optimal solution of the energy function. The local methods can achieve good matching results in disparity inconsistency regions, but the matching results are not robust. Global methods can achieve robust matching results, but it may cause fattening problems in edge regions. In order to take full advantages of local methods and global methods, this paper combins the two methods together and proposes a new image-guided non-local stereo matching method with three steps optimazition. Firstly, HOG feature has been improved so that the improved HOG feature is insensitive to the nonlinear radiation distortion, which is regarded as the cost metric in this paper; secondly, a new image-guided non-local cost aggregation stragety is designed, which avoids the limitation of local methods that similar pixels should have the same disparity, and it can achieve robust matching results in both edge regions and textureless regions; secondly, a semi-global matching method based on the aggregated cost is proposed, which can strengthen the cost aggregation in texture regions; Finally, a new image-guided disparity interpolation method is proposed, which is able to refine the disparity image.3) A dense matching method based on LiDAR point cloud constraints is proposed. Traditioanl matching methods with LiDAR constraints regard LiDAR points as reliable control points. Only few outliers are allowed in LiDAR point clouds. However, due to the occlusions, time differences and multiple reflections, there may be a large amount of outliers in local scope, which will influence the matching accuracy of traditional methods. This paper regard LiDAR points as a weak "soft constraint" and a strong "soft constraint". Firstly, LiDAR points are regarded as weak "soft constratins", and outliers are filtered based on disparity images after matching; then, the LiDAR points after filtering are regarded as a strong "soft constraint", and triangular meshes are constructed to guide the matching process.4) A new consistency check and fusion method for LiDAR points and dense matching points is proposed. The fusion of LiDAR points and dense matching points is helpful for recovering the 3D information in occluded regions. But the system errors in LiDAR point clouds and the mismatches in dense matching point clouds will influence the accuracy and visuality of fusion products. This paper proposes a new consistency check for LiDAR points and dense matching points for accurate fusion. Firstly, check LiDAR points based on dense matching points, and eliminate the LiDAR points which is higher than the dense matching points in the same plane position; Secondly, extract a series of small point clusters by fast elevation image segmentation, and check the small point clusters based on the LiDAR points. Small point clusters will be eliminated if there exist lower LiDAR points in the same plane position.In order to test the correctness and the reliability of the proposed method, various experimental data are used in the experiments of every key algorithms proposed in this paper, including testing for cost metrics, testing for the image-guided non-local cost aggregation stragety, testing for the dense matching method under Benchmark data and aerial images, testing for the LiDAR point filtering, testing for the dense matching based on LiDAR constraints, and testing for the consistency check of LiDAR points and dense matching points. Finally, adopt the proposed method for multi-view image matching with LiDAR point clouds, and test the effectiveness of the proposed method from the aspects of accuracy, point number, running time and so on. |
PDF Full Text Request