Elevation Extraction and Object Recognition from Aerial Stereo Images

Stereo image matching is to find conjugate points on overlapping images pair, i.e., search for the same image pixels of a ground point on the left and right images. When the image orientation parameters are known, the 3D (X,Y,Z) coordinates of the ground point can be calculated. This process is the core part of DEM (Digital Elevation Model) generation in digital photogrammetry. Research on this topic was very intensive in the past decades, and many different methods have been proposed. But, this problem is still not completely solved. The difficulty comes from many sources, such as radiometric distortion, geometric distortion, occlusion, repetitive pattern and lack of features, etc. [1]. The matching result usually has lots of mis-matches, especially along the boundaries of the ground objects, like road and house. In this paper, a method to use areas, multi-scale regions, recognized objects as matching entities is proposed. This method can dramatically reduce the mis-match, and generate more meaningful matching result. The presented research scheme includes the following key steps: (1). Multi-scale image segmentation (2). Object recognition/Classification (3). Matching using areas, multi-scale regions, and recognized objects (4). Bare earth surface reduction (5). Manual editing of the automated matching result The post processing steps (4) and (5) are needed to get the bare earth DEM, as the result of the automated stereo matching process is a DSM (Digital Surface Model).