Context and objectives
As the role of spatial information becomes more important (e.g. decision making and management for government bodies and companies, geomarketing, car navigation) pressure is put on the industry to produce detailed and up-to-date digital maps. Computer vision and object recognition can play an important role to improve the production of these maps, which is very time-consuming and price-absorbing.
One issue is the problem of making different data sets geometrically consistent. This is essential for a coherent use of data sets supplied by different data producers. The coming of more detailed spatial databases (on geometry, attribute and feature level) increases the problem of making the data sets consistent. The need for these high accuracy products is predicted for future markets (e.g. advanced driver assistance in car navigation, land management, etc.) and the availability of high resolution source material (e.g. Ikonos 1m resolution satellite imagery) makes computer vision technology accessible. This technology is useful not only to upgrade the database but also to detect and remedy areas with spatial quality problems in order to ultimately end up with a homogeneous European database.
Expected scientific results
Line detection proved to be a powerful technique for detection of roads, not only in fields as is often shown in the literature, but also in suburban and industrial zones. The technique is very generic and useful for a variety of sensors (satellite/aerial, visual/SAR). The detection is not perfect however. For registration more stable objects were needed than the road fragments that were detected. Crossroads proved to be a good choice. Detection was adequate and more important the number of false detection could be kept low.
Based on the detection of crossroads, a correspondence between the image and the database was found using graph matching techniques. The main factor here was a good definition of the constraints which define the desired solution. A good definition of constraints allows for a good performance of the system. The transformation that is found is a morphing transformation, where each point has a counterpart and local transformation of image patches can be performed. This is very flexible and allows for local deformations to exist. By putting bounds on the allowed transformation, outliers can be detected. This is of high importance for quality control where inconsistencies should be detected.
The results were presented to Tele Atlas and the technology was found to be useful to be considered for integration within their production environment.