Study and Implementation of an Object Tracing Algorithm for the Deflectrometric Shape Qualification of Parabolic Trough Collectors

Kurzfassung

This paper presents a comprehensive study of dierent segmentation methods to eectively and accurately extract the absorber tube reex from aerial images of parabolic trough collectors. It is part of QFly, currently running at the German Aerospace Centre (DLR), in which the slope deviations of a parabolic trough reector are determined based on airborne image acquisition. The advantages of this new shape qualication method primarily lie in the automated evaluation procedure. Other than alternative methods, QFly shall soon allow the determination of the geometrical accuracy of an entire parabolic trough eld providing essential information about the overall plant performance. The determination of slope deviations is based on the defectometric principle. The absorber tube reex, which is visible during image acquisition ights, yields enough information to geometrically reconstruct the reector shape. It is therefore of major importance to trace the reex locations as accurately as possible. In a former approach segmentation was done by means of a simple global otsu thresholding method with subsequent morphological cleaning. However, facing unfavourable contrast and lighting conditions, the method's limits have soon been revealed. The conducted study shall contribute to nd an alternative algorithm suitable for the QFly application with focus on automation, reliability, accuracy and CPU time. In a rst step, performance was enhanced by augmenting the segmentation procedure with pre- and post-processing features accounting for image-related phenomenons that typically cause segmentation algorithms to fail. An eective background reduction method as well as a bilateral lter have been suggested to encounter global discontinuities in intensity levels and to remove image noise respectively. Morphological cleaning as the former post-processing instrument was replaced by two strictly edge conservative methods. In this respect, especially the supervised classication method, implemented to remove sun reex patterns with absorber tube atypical characteristics, proved to be intrinsically more powerful than the previous approach. Pre- and post-processing algorithms have been meaningfully combined with global and local thresholding methods to subsequent segmentation cycles. All cycles have been applied to a sample of 32 images which were supposed to represent the worst case scenario. Resulting binary images have been compared with the corresponding manually generated ground-truth image. A validation method was introduced, which uses a distance weighting routine. The edge congruence is measured in EMM. Compared to the previous result, the validation showed notable improvement. The otsu method could reach the best results of all global candidates. The EMM was reduced by 46.96 % without post-processing and 61.06 % with post-processing. Processing took 1.62 s and 3.42 s longer than before. The highest overall accuracy was achieved by the local, fully automated Yan-Zhang-Kube method. The EMM dropped by 62.24 % if no post-processing was applied and 66.84 % with subsequent postprocessing. Additional CPU-Time was 43.28 s and 44.89 s respectivly. Accuracy and deviation of both global and local algorithms are still too high as to allow fully automated evaluations. Further improvement could be achieved by taking additional binarization methods into account. All approaches should operate on the colour domain instead of the photometric domain of a single colour channel. Based on the concept of partial truth dierent solutions from a whole set of segmentation cycles could be combined by means of a fuzzy logic. An extensive design parameter optimization is yet to be done.