AR GIS on a Physical Map based on Map Image Retrieval using LLAH Tracking Hideaki Uchiyama, Hideo Saito Keio University 3-14-1 Hiyoshi, Kohoku-ku 223-8522, Japan {uchiyama,saito}@hvrl.ics.keio.ac.jp Myriam Servi`eres, Guillaume Moreau Ecole Centrale de Nantes - CERMA IRSTV Rue de la No¨e, BP 92101, 44321 Nantes Cedex 3, France {myriam.servieres,guillaume.moreau}@ec-nantes.fr Abstract This paper presents a method for retrieving a cor- responding map of a captured map image from a map database. Our method is inspired from LLAH based Document Image Retrieval (DIR). LLAH is a method for recognizing a point by using a LLAH feature com- posed of its neighbor points. Since Map Image Retrieval (MIR) is achieved by analyzing distribution of intersec- tions, the LLAH feature is used in order to describe the distribution. In our method, registration and retrieval in LLAH based DIR are improved for reducing the com- putational costs of the retrieval. In addition, the LLAH features are updated while a camera is moving. Our improvements enable MIR to the case of strong camera tilting, occlusion and fewer intersections. 1 Introduction Geographical Information Systems (GIS) have be- come one of essential tools for handling urban develop- ment. GIS include spatial data such as buildings with its temporal changes. Compared with a paper map, digital GIS data can be updated anytime and be used for several applications. One of the research issues about GIS is the way of displaying the spatial and temporal GIS data. Previous works [6, 12, 13] have shown the advantages of Aug- mented Reality (AR) techniques to display GIS data on a physical map. For displaying the data at a precise position, these systems have to compute the geomet- rical relationship between a camera and a map. In these works, they have computed the relationship by using ARToolKit [6], local feature matching [12] and analysis of distribution of intersections [13]. Use of the distribution of intersections will be suitable for enhanc- ing general use of the system because the distribution should be the same in maps from different manufac- tures. However, there were some limitations for the camera motion in [13] because they applied only LLAH based Document Image Retrieval (DIR) [9], which is a method for retrieving the corresponding page of a cap- tured document image from a document database. In this paper, we propose AR GIS based on Map Im- age Retrieval (MIR) inspired from LLAH based DIR. Our method allows retrieval with fewer points with strong camera tilting and faster processing against LLAH based DIR. A captured map is retrieved from a map database by analyzing a positional relationship of intersections described by a LLAH feature. Since the positional relationship changes according to the changes of a user’s viewpoint, a new positional rela- tionship is registered to enable free camera moving, which is called LLAH Tracking. For evaluating our system, the minimum number of intersections for MIR will be discussed in the experimental results. From the aspect of computational costs, we will show that our algorithm will be compatible with AR systems. In ad- dition, we will show our system can work in the case of strong camera tilting and occlusions. The rest of the paper is organized as follows: we will present related works of object recognition meth- ods in Section 2. From the related works, the details of LLAH based DIR which inspired our method will be explained in Section 3. We will then provide an overview of our system in Section 4 and its algorithm including our improvement from LLAH based DIR in Section 5. Finally, experimental results for evaluating our algorithm will be presented in Section 6. 2 Related Works The research issue for finding a match with a query object using natural feature points has been addressed in various ways. The feature points can be described as a high-dimensional vector such as SIFT [8], SURF [2], SIFT and Fern [15]. The features are robust in terms of change of illumination, scale and rotation. The search method can be addressed as a nearest neighbor search- ing problem by the following approaches: ANN [1], locality-sensitive-hashing [4] and a vocabulary tree [10]. Rich descriptors are well suited to the matching of fea- ture points with few repetitive texture patterns. On the other hand, 2D maps can be presented in different ways without using descriptors. For example, there are roads, their connectivity and intersections, which should be same according to the manufacturers. Use of such topological features enhances general use of the system because it enables use of maps from different manufactures. Regarding the object recognition by topological fea- tures, geometric hashing (GH) is a 3D object recogni- tion method by using corners and edges [7]. Since a geometrical invariant in GH needs huge computational MVA2009 IAPR Conference on Machine Vision Applications, May 20-22, 2009, Yokohama, JAPAN 12-3 382