Automated Image-Based Verification Of Road Databases*

Transcription

1 Automated Image-Based Verification Of Road Databases* M. Gerke 1, Ch. Heipke 1, A. Busch 2 1 IPI - Institute of Photogrammetry and GeoInformation University of Hannover Nienburger Str. 1, Hannover, Germany {gerke,heipke}@ipi.uni-hannover.de 2 Bundesamt für Kartographie und Geodäsie Richard-Strauss-Allee 11, Frankfurt, Germany andreas.busch@bkg.bund.de SUMMARY In this abstract a system for automated road verification is introduced. This system is being developed in cooperation with the German Federal Agency Cartography and Geodesy (BKG). The background is to support a human operator in verifying a given set of road vectors (ATKIS DLMBasis). An up-to-date orthoimage is used for this verification. By means of image analysis algorithms a majority of correct objects should be identified, so that the operator just has to focus on the remaining roads. We give a short overview and present results of a 40km 2 testscene. Wrong decisions of the system are explained and analysed in more detail. The overall results show the potential of the approach. KEYWORDS: Database, GIS, Imagery, Networks, Parameters, Quality, Reliability INTRODUCTION The verification of existing geo-spatial information is an important task for Geographic Information Systems. Today, this work is carried out mainly manually by an operator, who compares vector data from databases with remotely sensed imagery or other information. A particularly important class of geospatial objects are roads. In this paper a system for automated road data verification using digital image processing for the extraction of roads from aerial imagery and topological analysis in order to optimise the whole process in terms of reliability and efficiency is presented. The main goal is to call the human operator s attention only to parts of the network where the automated process did not find sufficient evidence of a road. The road extraction is supported by the use of prior knowledge on the global level (whether the road is situated in rural, urban or forest ar 51 eas), and information on the road geometry and its attributes. This system is being developed in the framework of a project carried out with the BKG (German Federal Agency for Cartography and Geodesy). In Germany ATKIS (Authoritative Topographic Cartographic Information System) represents the official topographic reference dataset for the whole country. The dataset with the highest resolution is called DLMBasis. Its content approximately equals a topographic map of a scale of 1:25,000. The federal states of Germany are responsible for the acquisition of the DLMBasis, whereas the BKG produces a nationwide, homogeneous dataset by merging the datasets delivered by the 16 federal states. Therefore, the BKG is highly interested in an efficient and reliable determination of the quality of the delivered database objects. To organise the verification of the data independently from its capture, a semi-automatic working environment has been installed at BKG. * Abstract of a publication appeared in the ISPRS Journal of Photogrammetry and Remote Sensing (Gerke et al., 2004). 7 th AGILE Conference on Geographic Information Science 29 April-1May 2004, Heraklion, Greece Parallel Session 7.1- Remote Sensing II 569

2 Besides the obligatory GIS component this workstation includes an automatic image analysis module, which supports a human operator in the verification. The objects from the database are compared to an up-to-date orthoimage in order to achieve information on their quality. The main objects of interest are roads and area objects (like built-up areas and vegetation areas) as they are the objects changing most frequently. Automatic image operators being able to detect the objects of interest use as much prior knowledge as possible from the database, such as contextual and geometrical information. The whole automatic scene analysis comprising feature extraction, grouping of features and verification of the given ATKIS objects is managed by GeoAIDA (Geo Automatic Image Data Analyser, refer to (Bückner et al., 2002)). In our presentation we will focus on the verification of road objects, details on the verification of area objects can be found in Müller et al., SYSTEM DESIGN AND ROAD EXTRACTION APPROACH The system for automated road verification includes three modules: An Automatic Pre-Processing module, the Main Automatic Processing module and an Interactive Post-Processing module (c.f. Fig.1). In the Automatic Pre-Processing phase the GIS Component exports the road objects to be verified from the database, including their geometric descriptions and attributes, such as road width. Moreover, the knowledge about the global context contained in the ATKIS DLMBasis is obtained. Figure 1. System overview The Main Automatic module consists of the Process Control Component and the Image Analysis Component. The Process Control Component is a communication layer between the GIS and the Image Analysis Component: It makes the information from the database available for object extraction in an appropriate manner. The Image Analysis Component itself consists of the Verification and the Change acquisition module. Finally the GIS Component is used again to support interactive post editing. In the following we focus on the Verification module. 570

3 THE VERIFICATION MODULE The Verification module is designed to check whether the roads from the database keep a predefined positional accuracy as well as to detect commission errors (a road from the database does not exist in the reference orthoimage). In order to solve this task a region of interest is defined for each road object from the database, depending on its geometric description. More precisely, a buffer around the vector representing the road axis is defined, the buffer width complies with the corresponding attribute given in the ATKIS database, and additionally the nominal accuracy of the ATKIS road objects being +/- 3m is considered. If the road width fails a plausibility test or is not available at all, a predefined value is taken. This step is followed by the selection of an appropriate algorithm for a road extraction to be executed in the image domain of the buffer. In this system the road extraction algorithm as introduced by Wiedemann (Wiedemann, 2002; Wiedemann and Ebner, 2000) is currently used. This approach models roads as linear objects in aerial or satellite imagery with a resolution of about 1 to 2 m. The underlying line extractor is the one introduced by Steger (1998). The initially extracted lines are evaluated by fuzzy values according to attributes like length, straightness, constancy in width and constancy in gray values. The evaluation is followed by a fusion of lines originating from different channels. In this case panchromatic imagery is used, but the line extractor is applied twice: Firstly, using a bright line model (line is brighter than the background) and secondly using a dark line model (line is darker than the background). The reason is that according to a lot of experiments roads in images often fit to one of those models. The last step in road extraction as applied in the Verification module is the grouping of the single lines in order to derive a topologically correct and geometrically optimal path between seed points according to some predefined criteria. The decision if extracted and evaluated lines are grouped into one road object is made corresponding to a collinearity criterion (allowing a maximum gap length and a maximum direction difference). The described road extraction software was adapted to the given tasks, especially by applying individual parameters for the given context areas and the extraction for each road object separately. If the complete ATKIS object was reconstructed by this means, it gets accepted and rejected otherwise. As proposed in (Gerke et al., 2004) the road extraction algorithm is applied twice: In the first phase a strict parameter control (relating to contrast thresholds for the initial line extraction and gap length to be bridged automatically) is used. The road objects accepted in this step are called accepted I roads, the rejected ones are denoted rejected I roads. In the second phase a more tolerant set of parameters is chosen. The idea is to minimize the number of false positives in the first phase. This is important in order to ensure a good reliability of the verification. The selection of roads having been rejected in the first phase but to be checked again in the second phase is based on a topological investigation of the given network. It is assumed that already accepted road junctions are connected by the shortest possible path using the road objects from the database. All rejected I roads being situated one these shortest connections are chosen for the second phase. The thresholds for the contrast conditions and the gap length to be bridged automatically are adapted in order to obtain a more tolerant parameter control. After the second extraction phase each appropriate road object is evaluated again. Similarly to the first evaluation step the objects are denoted as accepted II if the road extraction algorithm could extract a valid road in the corresponding buffer, and as rejected II otherwise. In Fig. 2 an example is given: The left image shows an orthoimage, in the right image the accepted I roads are displayed in white, the accepted II roads in dashed-white and the rejected objects in black. As ATKIS objects lying in settlement areas are not considered, those regions are masked out. The finally accepted network is much denser than the road object being accepted in phase I. A more detailed analysis of the results is given in the next section. 571

4 Orthoimage Verification Result Figure 2. Example for the current approach RESULTS AND DISCUSSION A prototype of the system has been tested with 10 orthoimages covering an area 40 km 2 near Frankfurt/Main, Germany, and containing 2356 road objects in rural context. All roads together have a length of 374km. The panchromatic orthoimages have a ground sampling distance of 0.4m, but they have been resampled to 1.6m in order to meet the demands of the road extraction algorithm. The road extraction approach is not able to extract roads in urban areas, therefore these regions are not considered. The results from the verification where compared with a reference dataset in order to evaluate the system, refer to Tab 1.: For each accepted object we checked whether it is correct or not, and we did the same for the rejected objects. Ideally the elements beside the main diagonal should be zero. The values in the accepted/incorrect-field are an indicator for the reliability of our approach (False Positive FP: The smaller the value, the more reliable are the results). A small value for FP is crucial for our approach, as we assume that all objects which are accepted by the system do not need to be checked by a human operator. Thus, FP errors cannot be detected at a later stage. The obtained result of only 1,5% FP errors is very encouraging, nevertheless, the FP errors are further analysed in the following. The number rejected/correct roads (False Negative FN) should of course also be small, however this value is not as important as the first one, because such roads are subsequently checked by the human operator.. Thus, FN is an indicator of the efficiency of the system. In our test we reached a value of 27.2%.Together with the correctly rejected roads (2.4%) this means that compared to a completely manual procedure only about 1/3 of the roads needs to be analysed interactively 572

5 Accepted by the system Rejected by the system Correct objects 68,9% 27,2% Incorrect objects 1,5% 2,4% Table 1. Evaluation of road object verification ANALYSIS OF FALSE POSITIVE ERRORS Label errors, i.e. road objects from the database belonging to another object class in reality (such as river) did not appear in the dataset. Therefore, all FP-errors are related to cases where the geometrical accuracy is not maintained by the ATKIS object or where no road could be found in the image at all. In Table 2 an overview of the False Positives is given. Image Sum No. 13 (4,5%) 0 3 (1,7%) 0 3 (1%) 1 (0,5%) 7 3,8%) Table 2. Overview of False Positives 9 (4,0%) (1,5%) The FP-errors are divided into four classes: 1) The ATKIS road object is not situated inside the predefined tolerance of +/- 3m, but the system has accepted the object as it was found inside the buffer; 2) The ATKIS road object is very short and wrong, but the system has found a valid road; 3) Interfering objects like trees or fences obtain a good evaluation score and are therefore used for the road extraction; 4) The wrong ATKIS road object lies parallel to plough furrows or path of tractors, respectively; these are extracted. The numbers for this classification are given in Table 3. As can be seen the first group is the most important one. In the following, examples for all four groups are presented and discussed. Type of error ATKIS object outside tolerance ATKIS object very short Interfering objects Plough furrow/ path of tractor Number of objects Table 3. Absolute number of classified FP-errors The first kind of errors (ATKIS object outside tolerance but accepted) are related to the fact that the ATKIS DLMBasis contains the centreline of the roads and has a nominal accuracy of +/- 3m. The buffer for the automatic line extraction algorithm must contain the whole potential road, i.e. +/- (3m * road width) on each side of the centreline. Inside this buffer often some parallel line-shaped objects like ditches or shoulders are situated which are extracted as lines. As a differentiation between the real road centreline and such parallel lines is currently not implemented, the system accepts all extracted roads in this buffer. In other words, the current system can not check whether the nominal accuracy of +/- 3m is maintained; and thus objects which are found inside the buffer with a radius of +/- (3m * road width) are accepted. In Fig. 3 such a situation is shown: The upper left image shows the orthoimage, the upper right shows the same image with the ATKIS object superimposed in black. Obviously the geometric accuracy of +/- 3m is not maintained in this case, but the buffer being broader than this value contains the road (lower left image). The result of the road extraction is given in the lower right image. As the extracted road connects the first and last vertex of the road from the ATKIS DLMBasis this object is accepted. 573

6 Figure 3. Example of False Positives due to the buffer width being too large The second class of FP-errors (acceptance of short lines, though not correct) is related to the fact that the search for a connection between the first and the last vertex of the ATKIS road object (grouping of extracted lines) is supported by inserting a small vector (seed vector) pointing from the end vertices towards the respective next one. If no line between these vertices from the ATKIS object was extracted, the object is rejected. But, if small line segments are found in this area and the resulting gap between such segments and the artificial seed vectors from the ATKIS object is short enough for an automatic closing, this object is accepted. Typically, this error occurs mostly with the tolerant parameter control as the length of the maximum allowed gap which is bridged automatically is relatively small. In Fig. 4 an example is given: The upper left image shows the orthoimage, in the upper right image the ATKIS road object is shown, which does not maintain the geometrical accuracy of +/- 3m. The lower left image shows 574

7 that a lot of small line segments have been extracted inside the buffer. As the gap between these segments is small enough to be bridged automatically, the object was successfully extracted (lower right image) and therefore the ATKIS object was accepted. This class of FP-errors more often occurs with a tolerant parameter control. It is obvious that in this case the strategy followed in the topology supported verification helps to reduce the number of FP-errors: Due to the fact that those short objects are often situated at the border of the image (road objects are clipped), they won t be chosen for phase II. Figure 4. Example of False Positives due to the problem with short objects The third class of FP-errors can be explained by the inability of the used road extraction model to model local context, i.e. the interrelationship between single objects. Objects like rows of trees or fences are not explicitly modelled and therefore may be incorporated into the grouping process, if the line extraction was successful. Again, an example is given (refer to Fig. 5). The upper left image shows the orthoimage and on the upper right image the original ATKIS object is shown. In the southern part the ATKIS object runs across the row of trees and does not maintain the maximum geometric tolerance. The line extraction algorithm finds lines on and beside the trees (lower left image), which are connected in the grouping process (lower right image). In general the extraction of row of trees supports the extraction of roads, as these are often placed besides each other (c.f. e.g. (Butenuth et al., 2003)). In this case the missing differentiation between these two object-classes in conjunction with the buffer width (see above) leads to an acceptance of the incorrect road. The FP-errors assigned to this class also occur more often with the tolerant parameter control. 575

8 Figure 5. Example of False Positives due to lack of local context modelling The fourth and last type of FP-errors is related to the fact, that sometimes the appearance of plough furrows or paths of tractors are similar to small roads (paths) between fields. If such structures are covered by the buffer around the ATKIS object and if their contrast values fit to the given control parameters, these paths are extracted as roads. Figure 6 gives an example: The ATKIS road has an offset larger than +/- 3m regarding the road which can be seen in the image (upper right). The buffer of this object covers three small paths in the field, which have been extracted (lower left). Finally a road could be extracted, thus leading to a false positive error. These FP-errors are a general problem of the approach if low contrast thresholds are applied: On a field such structures are extracted as lines and as there are in general no disturbances like trees the extraction result is likely to be labelled a road. 576

9 Figure 6. Example of False Positives due to the extraction of plough furrows / path of tractor ANALYSIS OF FALSE NEGATIVE ERRORS In Table 4 statistics for the false negatives (correct, but rejected) are given. An analysis of these objects being rejected though being correct shows that several causes may lead to the rejection: a) The contrast is very weak, b) Lines are not extracted due to the missing step-model implementation, c) Roads are hidden by dense rows of trees and these rows have not been detected as lines. Often, parts of roads are also not extracted due to one of the above reasons. In these cases the resulting gap is too large to be bridged automatically, even in the tolerant second examination phase. Image Sum No. 84 (29%) 56 (26%) 38 (22%) 73 (37%) 65 (22%) 73 (37%) 42 (23%) 40 (17%) Table 4. Overview of False Negatives 81 (27%) 90 (40%) 642 (27%) 577

10 CONCLUSIONS In this paper a two-step graph-supported approach for road verification from aerial imagery is introduced and evaluated. The background of this work is to design a system which supports a human operator in verifying an existing road database. The first phase focuses on reliability. In the subsequent phase II the efficiency may be increased, keeping in mind that the number of false positives, for example due to misinterpreted paths on fields becomes higher. This example shows a conceptual limitation of the system: As the verification is done per object, similar neighbouring objects are not detected and thus field structures are possibly misinterpreted as road. Another limitation concerns the examination of the positional accuracy: Independently from the used parameter control the used approach may cause FP-errors if a road object exceeds the nominal positional accuracy of +/- 3m but can be found inside the buffer considering the width of the road. But if we leave these restrictions out of consideration as its effect on the overall result is rather small the system helps to significantly reduce the amount of time for the quality control: In the exemplary investigation the number of False Negative Errors is less than 30%, i.e. more than 70% of the time for manual verification is saved. Another conclusion is that in this small spot check the quality of the ATKIS road database is very good: Just about 4% of the objects are incorrect. Therefore one may ask, why this detailed evaluation of the verification approach is necessary. One answer to this question is that quality control is necessary in any case. Another one is that the knowledge from the verification step will support the detection of new roads: Supposed that a majority of the new roads are connected to the existing network, the search space for new junctions is reduced and already extracted and evaluated lines can be used as seed objects for the extraction of connecting roads. Furthermore it has been observed that often the geometry and the topology of the ATKIS road network is correct, but the absolute position may be improved (compared to its position in the orthoimage). Therefore one may take into account the possibility to keep the geometry and the topology of the object from the database but shift its position towards the extracted line. BIBLIOGRAPHY Bückner, J., Müller, S., Pahl, M. and O. Stahlhut, Semantic Interpretation of Remote Sensing Data. In: International Archives of Photogrammetry and Remote Sensing, Com.III, Graz, Vol.XXXIV Part B3a, pp Butenuth, M., Straub, B.-M., Heipke, C. and F. Willrich., Tree Supported Road Extraction from Arial Images Using Global and Local Context Knowledge. ICVS 2003, Graz, Editors: Crowley, Piater, Vicze, Paletta, Lecture Notes in Computer Science 2626, Springer, pp Gerke M., Butenuth M., Heipke C., Willrich F., 2004: Graph-supported verification of road databases, Journal of the International Society for Photogrammetry and Remote Sensing, Vol. 58, No. 3/4, pp Müller, S., Weis, M., Liedtke, C.-E., and M.Pahl.,2003. Automatic Quality Surveillance of GIS Data with GeoAIDA. In: International Archives of Photogrammetry and Remote Sensing, Munich, Vol.XXXIV Part 3/W8, pp Steger, C., An unbiased detector of curvilinear structures. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 20 (2), pp Wiedemann, C., Extraktion von Straßennetzen aus optischen Satellitenbildern. DGK bei der bayrischen Akademie der Wissenschaften, Reihe C Dissertationen, Heft 551, München. Wiedemann, C. and H. Ebner., Automatic Completion and Evaluation of Road Networks. In: International Archives of Photogrammetry and Remote Sensing, Amsterdam, Vol. XXXIII, Part B3, pp