pomiar pęknieć w betonie tylko bez hierarchicznych, reszta
[ Pobierz całość w formacie PDF ]
//-->Photogrammetric Record,17(99): 453–464 (April 2002)AN OPERATIONAL APPLICATION OFAUTOMATIC FEATURE EXTRACTION:THE MEASUREMENT OF CRACKS INCONCRETE STRUCTURESP M. D(pmdare@unimelb.edu.au)H B. H(hanley@sunrise.sli.unimelb.edu.au)C S. F(c.fraser@eng.unimelb.edu.au)University of Melbourne, AustraliaBR(b.riedel@tu-bs.de)W N(w.niemeier@tu-bs.de)Technical University of Braunschweig, GermanyAbstractAn understanding of the evolution of cracks in concrete structures dueto long term natural deformation is important to civil engineers, but quanti-tative measurements can be difficult to make. However, digital imaging offersa potential solution. This short paper illustrates the operational applicationof automated image processing techniques for accurate, multi-temporal crackmeasurements. The first part of this paper provides an overview of automaticfeature extraction, essential for automatic crack detection. The latter partdescribes the methods developed for detecting and measuring cracks. Due tothe long term nature of the application, operational results have yet to befinalised, although sample results are presented.K:automation, crack monitoring, feature extractionIFhas been the cornerstone of many applications in photogram-metry and remote sensing for several years (Forstner, 1993). The inevitable trend¨towards automation in all things digital has meant that research into automaticfeature extraction has received considerable attention for some time now. Althoughsignificant progress has been made in many areas of automatic feature extraction(for example, Firestone et al., 1996; Sowmya and Trinder, 2000), transfer of thosealgorithms from the research community to the commercial domain has been slow.In the photogrammetry and remote sensing research communities, automaticfeature extraction is being used for a considerable number of different applications.Photogrammetric Record,17(99), 2002453Det al. Operational application of automatic feature extraction: cracks in concreteFor example, ‘‘high level’’ processes include the extraction of cartographic featuresfrom digital images, and building extraction and reconstruction for developmentof three dimensional city models (McKeown et al., 1999; Gulch, 2000). However,¨these are essentially topics of research which have yet to be exploited commercially.It is the ‘‘lower level’’ processes that have been successfully commercialised.Automatic detection of fiducial marks (Schenk, 2000) has been incorporated intomany digital photogrammetric workstations, and the automatic detection of targetshas been utilised in close range photogrammetric systems (Fraser, 1997; Ganci andHanley, 1998).As digital imaging systems become more widespread, demand for usable highlevel automatic feature extraction algorithms (as opposed to research tools) willincrease. One example application is the automatic extraction of crack features inconcrete structures from close range digital imagery. In this paper the issues associ-ated with developing application-defined automatic feature extraction algorithmsare exemplified by reference to this problem.A F EBackgroundWhen discussing automatic feature extraction from digital images it is import-ant to define exactly what is meant by this phrase. In the scope of this paper, andto some extent the wider photogrammetric community, a feature can be thoughtof as a spatially connected ensemble of pixels in image space which represents thephysical extent of an object in object space. Extraction of features refers to isolatingthose pixels which represent a particular object from all other pixels in the image.The degree of automation specifies what level of human interaction is required inorder to extract the features successfully.A significant amount of research has been carried out, and continues to becarried out, on the subject of automatic feature extraction. A search of the Institutefor Scientific InformationScience Citation Index(ISI, 2000) found over 40 journalpapers, published between 1990 and 2000, that had the phrase ‘‘automatic featureextraction’’ in their title. That value increases to almost 600 when the search isexpanded to allow the search terms to be part of the keyword field or abstract.(Note that these ISI searches did not take into account papers published in theInternational Archives of Photogrammetry and Remote Sensing,which would haveincreased the results significantly.) Furthermore, six papers related to automaticfeature extraction were published in thePhotogrammetric Recordin 2000 alone,and a similar number over the preceding three years. Finally, the Ascona work-shops, entitledAutomatic extraction of man-made objects from aerial and spaceimages(Gruen et al., 1995; Gruen et al., 1997), included 65 papers on either auto-matic or semi-automatic feature extraction. These figures indicate without doubtthat automatic feature extraction is currently a major research topic.It is convenient to subdivide this research into two categories: investigation ofnew feature extraction techniques, and implementation of current techniques tosolve specific problems. The subject of automatic crack detection, described in theremainder of this paper, concentrates on the latter of these two categories. Therefore454Photogrammetric Record,17(99), 2002Det al. Operational application of automatic feature extraction: cracks in concretethe two issues of most relevance are the practical realisation of automatic featureextraction, and automation itself.Practical Realisation of Automatic Feature ExtractionThe transfer of automatic feature extraction algorithms from ‘‘proof-of-concept’’ research projects to commercial products has been slow. Some attempthas been made to market stand-alone automatic feature extraction software pack-ages, but these systems are still few in number and have not yet enjoyed greatcommercial success. The reasons for this are varied and may be specific to individualsoftware packages, but one point should be emphasised: feature extraction is appli-cation dependent. For example, automatic algorithms for extracting buildings fromdigital imagery are quite different from road network extraction algorithms (com-pare Gruen and Li (1995) with Gulch (2000)). Furthermore, algorithms for extrac-¨tion of roads from satellite imagery are different from those for extracting roadsfrom aerial photographs (compare Karathanassi et al. (1999) with Laptev et al.(2000)). Consequently, it is not feasible to attempt to develop generic feature extrac-tion algorithms (those that can extract any type of feature from any type of image),or even object-specific feature extraction algorithms (those that can extract aspecific class of features from a range of image types). Therefore, commercialfeature extraction software must incorporate a wide range of algorithms in orderfor the software to be widely applicable—something that is not necessarily commer-cially viable, especially considering the rather limited size of the potential market.The Concept of AutomationNo feature extraction algorithm can claim to be truly fully automatic. Thereis always a need for human intervention at some point in the processing chain,even if only to set the algorithm running on the correct image. Therefore, whendeveloping an automatic feature extraction algorithm it is necessary to define whatis an acceptable level of human interaction. It is proposed that the selection of thisdefinition should be based upon three criteria: accuracy, efficiency and repeatability.Accuracy and automation are intrinsically linked, but the relationship betweenthe two is not always clear-cut. For example, the accuracy at which a humanoperator can locate tie points in a pair of aerial photographs is of the order of3mm,whereas automatic image matching techniques (which rely on automaticextraction of similar point features) achieve approximately 0·1 to 0·4 pixels.However, the automatic technique is able to extract so many more tie points thana human operator could achieve in a reasonable amount of time that, statistically,the automatic technique will provide a comparable final result.Efficiency is the principal reason for introducing automation into a processingsystem. Automating mundane or laborious procedures, which previously requiredsubstantial human intervention, will lead to significant savings in processing time,which, provided that accuracy is not compromised, is a very desirable result.However, automation does not necessarily imply efficiency. Automating high levelprocessing tasks which are trivial for a human operator to perform, but consider-ably more difficult for an automatic algorithm (such as ensuring that an aerialPhotogrammetric Record,17(99), 2002455Det al. Operational application of automatic feature extraction: cracks in concretephotograph has been correctly exposed) can increase processing time and introducenew uncertainty into the results.Repeatability is fundamental to the automation of feature extraction algo-rithms. If an automatic algorithm generates high quality results for one specificimage, but cannot be used with any other images, then the operational utility ofthat algorithm is severely limited. The algorithm has to be able to provide qualityresults for any number of images that fall within the scope of the algorithm, evenif those images form a small and uniquely defined set. It is this ingredient that ismissing from many automatic feature extraction algorithms in the research domain,but is essential to algorithms in the commercial world.Crack DetectionBased on these principles, a procedure for automatic crack detection wasdeveloped. The goal was to implement a system which could measure the width ofa crack at regular spacings along its length. The crack would be digitally imagedat intervals over a long period of time, whilst forces are applied to the concretestructure. Therefore, by measuring crack widths from a sequence of multi-temporalimages, changes in the crack could be monitored. The remainder of this paperdescribes the development and implementation of the resulting system.A C DMD IBackgroundThe application of automatic feature extraction described in the remainder ofthis paper concerns the extraction of cracks from images of concrete structures,acquired with a digital camera. A series of images was acquired over a period oftime in order to monitor and measure changes in the cracks that occurred as aresult of long term natural deformation. Images were acquired using a JAI M1progressive scan camera with a 4·8 mm lens, positioned 0·6 m from the concrete.The orientation of the camera with respect to the target was maintained for eachimage to minimise geometric differences between successive images. A typical imageof a crack some 10 cm long is shown in Fig. 1.Automatic Crack DetectionThe image in Fig. 1 does not need to be examined in minute detail to realisethat fully automatic extraction of the crack feature is an ill-posed problem. Inparticular, the beginning and the end of the crack are ill-defined, and poor imageresolution and branches in the crack mean that its location is ambiguous.Furthermore, other features within the image that have similar radiometric proper-ties to the crack have the potential to confuse an automatic crack extractionalgorithm.This example of feature extraction is a typical case of where human inter-vention can greatly improve the efficiency of the extraction algorithm, rather thanhinder it. Identification of the crack in a cluttered image such as this is a simpleprocess for a human operator to perform, but an extremely difficult process for anautomatic algorithm. Therefore, an extraction algorithm was developed which456Photogrammetric Record,17(99), 2002Det al. Operational application of automatic feature extraction: cracks in concreteF. 1.Image of a crack in a concrete structure.relied upon minimal, but essential, human intervention. While in principle a fullyautomated procedure would be desirable, and, given appropriate image quality, itmight be possible to find the edges of the concrete structure automatically and thento search for any cracks on the surface using an edge extraction routine, in practicethis has proved impossible to achieve. In the specific application of crack detection,tuning of parameters for fully automatic feature extraction algorithms cannot beavoided, for the reasons described above, thus resulting in a loss of systemrobustness.Before a final extraction procedure was selected, two slightly different algo-rithms were tested: theroute-finderand thefly-fisher.The purpose of these twoalgorithms is to delineate the crack as a polyline. Once a polyline is defined,measurements of the crack width along this polyline are possible.The Route-finder AlgorithmThe route-finder algorithm relies on the human operator interactively selectingtwo points on the crack with a mouse-controlled pointer. A straight line is drawnbetween the two points which is accepted as the first (but obviously rather poor)approximation of the crack (the dashed line in Fig. 2).The algorithm moves a pre-defined distance along the line (selected by theuser) before a search is made for another point on the crack. This is done by takingthe profile of the pixel values perpendicular to the line. In this context, a profile isdefined as an array of pixel values, lying on a straight line of finite length. Thepixel values are sampled at regular intervals along the line. Neither the start point,the end point nor the intermediate points need be at integer pixel coordinates; pixelvalues that are off the regular pixel grid are determined using bi-linear interpolation.The point with the lowest pixel value in this profile is taken as being a point in thecrack (Fig. 3(a)). Obviously the assumption is made that the crack is a dark featurePhotogrammetric Record,17(99), 2002457
[ Pobierz całość w formacie PDF ]