In recent years, there has been considerable interest in self-driving cars and unmanned vehicles. One of the most important capabilities of self-driving cars is localization, which is the process of finding the position of an object on a given map. The Global Positioning System (GPS) is widely used for localization and navigation, but it is blocked and reflected by skyscrapers in an urban environment. Furthermore, dead-reckoning based on inertial sensors causes drift error, so other sensors, such as cameras and lidars have commonly been employed for urban navigation. These sensors are utilized to obtain information about the surrounding environment, and vehicle localization is conducted by comparing sensor measurements with given maps. We have selected a lidar in order to detect the building outline and match it with the aerial map. Different from other studies that employed road lanes and signs as features for matching, this enables localization robust to road conditions, weather, and illumination.
First, the semantic segmentation of an aerial image is performed using a neural network to detect the building outline. Most aerial maps contain a perspective projection distortion when describing a tall building, so we are proposing a correction method. The building outline is extracted from the corrected map, and the mutual information between lidar measurements and the building outline is obtained to measure the similarity between them. Finally, a particle filter framework is employed to localize the vehicle and match the map by using the mutual information as the weight of a particle. An experimental dataset is then used to validate the feasibility of the proposed method.