Matching Buildings Between 2.5D Maps and Dash Cam Images for Building Identification

Abstract

Technology that matches buildings between dash cam images and digital maps (GPS) assists in the identification of buildings. The identification enables to collect city information as well as the textures for 3D building models. However, the matching technology has two challenges. First, GPS locations can be highly inaccurate in cities that have tall buildings, which means the GPS sometimes needs to be relocated to capture the correct building features for matching. Second, it can be tricky to set the position and orientation of the camera by making correspondence of certain features of buildings in images and maps. In this work, we propose a method of matching buildings in dash cam images and 2.5D maps that uses the height information of the buildings equivalent to the LoD1 in the CityGML format. For the first challenge, we relocated GPS locations by using a map-matching method. For the second challenge, we adjust positions and orientations by matching the building edges in the images and maps after extracting the edges with a deep-learning-based detection model. Tests using real-world datasets demonstrated that our proposed method matched the buildings better than the baseline.