This thesis investigates the topic of automatic image matching with focus on the generation of Digital Surface Models (DSMs) by using imagery acquired from airborne linear array CCD sensors. The research has been motivated by the recent developments in photogrammetric equipment related to sensor technology, which have introduced a new field of research. Airborne digital cameras that employ linear array CCD sensors, exhibit different radiometric and geometric characteristics compared to conventional film based cameras and new methods are required for processing the data from these sensors and generating products for different applications. In addition, with respect to matching,existing algorithms are geared towards frame imagery, plus the degree of automation is limited. In most cases, current commercial Systems show poor success rate and require manual interaction for editing the matching results.
This study is embedded in the framework of the AIM project (Adaptive Image Matching), in which data from the airborne digital camera (ADS40) of LGGM (Leica Geosystems GIS & Mapping) are used as input. Existing matching algorithms are analyzed, further modified and new ones are developed aiming at integrating information from the special characteristics of the sensor in the matching philosophy. First, the camera architecture, the radiometric and geometric properties of the sensor, the calibration of the system, the ground processing Workflow and the sensor model are investigated.Then, this research is focused on two main issues, namely the evaluation and enhancement of the image quality and the development of the matching strategy. The radiometric analysis and the preprocessing part include methods for noise estimation, noise reduction, contrast and edge enhancement, radiometric balancing, reduction to 8-bit and processing of multispectral Channels. This part is important and if omitted, the matching Performance is influenced. The matching strategy consists of different modules that are evaluated individually but also on an integrated basis. In a nutshell, the individual matching aspects that are investigated are: the implementation of geometrical constraints,the derivation of approximate values, the extraction of features, the Integration of different matching methods and the quality control and error detection. Geometrical constraints are used to strengthen matching and are employed by means of quasi-epipolarlines. Due to the complex geometry of the images (position and attitude information for each line), epipolar lines do not really exist and the epipolar trajectory is modelled over a short length by a second or first degree polynomial equation. Moreover,hierarchical techniques are utilized to gradually refine the matching results. The investigations are performed with respect to generation of image pyramids, for which different filters are utilized and evaluated, and to the selection of the doublets as an optimal strategy towards better time Performance and reduction of propagation of matching errors to lower levels. Apart from the above, in the matching strategy feature- and arca-based methods, plus methods that aim at higher reliability and/or at precision, are combined and different primitives (grid points, edgels and edges) are used. The selection of edges, as matching entities, resulted from the evaluation of different feature extraction algorithms (points and edges), based on a set of criteria. Then, the efforts have been mainly focused on edge matching,in order to improve modellingof discontinuities. Different approaches were investigated that led to significant improvements: the use of height and continuity constraints for contour points and extensions of LSM for edge features. In the LSM for edge features both edgels and long, straight edges are handled. Moreover,the ADS40 with the configuration of the Channels on the focal plane and their viewing angles permits the use of more than one template and to facilitate the identification of errors oecurring in matching,especiallyocclusions. The role of the different combinations of Channels in matching is discussed and the matching block, based either on a single- or multi-template strategy, is described.
Other major aspects of these investigations are the quality control and error detection strategy. Each individual ray is checked based on a set of criteria and pre-defined error types. In the quality control, measures derived from different matching methods (multi-patch matching,LSM, edge matching) are combined, problematicrays are exeluded and each 3D point is computedfrom the good rays only. The Performance of the system has been evaluated over different areas of land cover and for different point classes (breaklines and points, on the ground and on anthropogenicobjects). A detailed analysis of the results and the Statistical measures that have been derived from the tests are presented and discussed. The derived accuracy of the automatic measurements is close to the accuracy of the manual measurements, According to the studies, blunders in the results of AIM are significantly less comparedto the results of the commercial system Socet Set 4.4.1 (SS). For AIM the matching accuracy on anthropogenic objects was 0.5-0.66 m, whereas for SS it was > 1 m, especiallyin dense built areas.