18-Geom-A4 Photogrammetry · May 2014
Nivaar worked solution (AI-drafted; not reviewed by a licensed engineer)
Paper format: National Exams, May 2014 — 3 hours, closed book (any non-communicating calculator permitted). SEVEN questions constitute a complete paper: Part A answer all of #1–#5, Part B answer one of #6/#7, Part C answer one of #8/#9. Marks are shown in brackets. All nine questions (including both alternatives in Parts B and C) are solved below for completeness.
Reference texts: Wolf, Dewitt & Wilkinson, Elements of Photogrammetry with Applications in GIS (4th ed., McGraw-Hill, 2014); Mikhail, Bethel & McGlone, Introduction to Modern Photogrammetry (Wiley, 2001); Kraus, Photogrammetry: Geometry from Images and Laser Scans (2nd ed., de Gruyter, 2007); Ghilani & Wolf, Elementary Surveying (15th ed.). Canadian mapping practice (NRCan / Canadian Geodetic Survey) throughout.
Question text not reproduced: the examination questions are © Engineers and Geoscientists BC. Open the official past paper (linked at the top of this page) to read the question, then follow the worked solution below.
7.1 What image matching is, and where it is used. Image matching is the automatic identification of conjugate (corresponding) points — the images of the same physical object point — in two or more overlapping digital images. It is the digital replacement for the operator who once placed a floating mark on conjugate detail, and it is the enabling step for most automated photogrammetric products. It is used in: relative orientation and automatic aerotriangulation (measuring tie points across a block); automatic DEM/DSM generation by dense stereo matching; digital orthophoto and true-orthophoto production (which need the surface model that matching produces); point-cloud generation in dense image matching / structure-from-motion; and feature tracking, mosaicking and change detection. In short, wherever conjugate points or a surface must be found without manual pointing, image matching does it.
7.2 Two methods of image matching.
(a) Area-based matching (ABM) — cross-correlation / least-squares matching. A small window of grey values around a point in the reference image is compared with candidate windows in the search image, and the position of best similarity is taken as the conjugate point. Similarity is measured by the normalised cross-correlation coefficient (peak correlation) or, more precisely, by least-squares matching, which also solves for radiometric (brightness/contrast) and geometric (shift, scale, shear) transformation parameters of the window to reach sub-pixel accuracy. ABM is very precise on well-textured surfaces but fails on repetitive patterns, low-texture areas, occlusions and large perspective/relief differences.
(b) Feature-based matching (FBM). Distinct features — interest points, corners, edges or blobs — are first extracted in each image with an operator (e.g. Förstner, Harris, or scale/rotation-invariant detectors such as SIFT/SURF), each feature is described by an attribute/descriptor vector, and features are then matched by comparing descriptors under geometric constraints (e.g. the epipolar line, or robust estimation with RANSAC). FBM is far more tolerant of scale, rotation and illumination differences and of large baselines, so it dominates automatic tie-point measurement and structure-from-motion; its matches are typically at feature locations rather than on a regular grid. In practice the two are combined: FBM gives robust approximate correspondences, then ABM/least-squares refines them to sub-pixel precision. (Relational/symbolic matching, which compares the structural relationships between features, is a third, higher-level approach.)
