Photogrammetric point clouds generation from high-resolution imagery in urban areas

Abstract

The advantages of high quality imaging, short revisit time, and lower cost make high-resolution imagery an attractive option for surveying and mapping applications. Automated reliable and dense image matching is essential for satisfactory achievement of photogrammetric 3D data derivation. Such matching, in urban areas, however, is particularly difficult, owing to the complexity of urban textures and severe occlusion problems on the images caused by tall buildings. Most of the traditional digital photogrammetry systems require intensive human interaction to process urban images. This research aims to exploit high-resolution imagery for 3D reconstruction in urban areas, and thus presents innovative approaches for image-based generation of 3D point clouds, emphasizing the reliable and dense image matching in urban areas. The proposed approaches consist essentially of several mutually connected components: the integrated feature point and edge matching, the integrated image segmentation and matching, and the dense matching propagation. Each is designed to solve those specific problems related to image matching in urban areas. Firstly, the integrated feature point and edge matching is designed to extract the feature points and edges, and obtain robust feature matches. The method consists of the following two steps, seed matching and feature matching. The seed matching aims to automatically obtain a small number of robust seed matches. The feature matching search for matches among the extracted features on the stereo image pair based on the self-adaptive triangulation constraint and other constraints. The output from the integrated feature point and edge matching tends to be a few but robust feature point and edge matches on the urban images. Secondly, an integrated image segmentation and matching strategy is designed to solve the occlusion problems on urban images caused by tall buildings for improving matching performance. First, image segmentation is conducted based on previous edge matching results and region growing techniques. The image segmentation results are then incorporated in the image matching process for occlusion filtering and segment-adaptive similarity measurement, which will help to improve the matching reliability and the avoidance of decision conflict in the disparity discontinuous areas. The segment-based occlusion filtering and segment-adaptive similarity measurement are to be employed in the matching procedure of the remaining extracted features and the dense matching propagation in later phase. Thirdly, effective dense matching propagation strategies are developed to obtain dense matches in urban images, including the local and regional propagation strategies. The dense matching will start from the previously obtained feature matches and the matching will be propagated from them to their neighbors based on the segment-based occlusion filtering, segment-adaptive similarity measurement, and other constraints, until all pixels in the images have been examined. From the obtained matching results, 3D point clouds are able to be generated through photogrammetric space intersection using the image orientation parameters. Experiment analysis using four sets of stereo images each representing different urban types are carried out. The accuracy of the 3D point clouds generated from the proposed approach is evaluated based on the comparison with the airborne LiDAR data. Comparisons are also made using the point clouds generated from the SGM (Semi-Global Matching) method. As evidenced by the experiment results, the proposed approach can generate not only the general geomorphological features of the surface, but also detailed features. The results from the quantitative accuracy comparison indicate that the proposed approach is able to generate 3D point clouds with a geometric accuracy comparable to LiDAR data, but offers a much higher point density. The proposed approach has similar performance with SGM in relatively flat areas, but shows a superiority to SGM in densely built-up areas. The developed approaches and methods presented in this research offer a viable alternative solution for 3D surface reconstruction in urban areas, and have the promise of facilitating the use of 3D data for other urban applications, such as urban planning, digital city construction, and smart city development.