Circular coding: A technique for visual localization in urban areas

Abstract

The task of visual localization in urban areas is to estimate the location of a camera on a given map by analyzing a street-view image photoed by this camera. Existing methods to accomplish this task perform either a matching between this street-view image and those ones with location tags stored in a database, or a retrieval of structural features of buildings extracted from this street-view image and ‘seen’ from the map. However, the former could fail when there is no database image that has high similarity to the query one, and the latter is generally of high complexity since the retrieval of geometric building information is time-consuming. In this paper, an efficient system for visual localization is developed, which is of the latter category but can achieve real-time performance while delivering high localization accuracy. The core of our developed system is a novel technique termed the circular coding, which is proposed to encode both the street-view image and the map locations with a uniform format, called the coding circle. With this technique, the same location will have the same encoding result in both coding cases, and thus the localization task is simplified to a retrieval process on coding circles that can be efficiently implemented at a low computational expense. Firstly, for each location on the map, the corners of building outlines are projected onto a unit circle to generate a database coding circle. As a result, a coding circle database that consists of all the generated coding circles is constructed. Then, for the camera location, the vertical corner lines (VCLs) of buildings are extracted from the street-view image and also projected onto a unit circle to generate the query coding circle. Finally, owing to a correspondence between the building corners and VCLs, the query location is estimated via a retrieval of its coding circle from the coding circle database. Experimental results show that our localization system is of fast speed and high accuracy, and thus can be used in various applications.