Camera-based vehicle localization at intersections using detailed digital maps

In recent years, a special research focus of the automotive industry was set on vehicle positioning. Though the accuracy of current simple positioning systems is sufficient for navigation devices, future technologies will be much more demanding. Examples for this future technologies are cooperative systems or enhanced map based assistance systems. All positioning system which will be elaborated always have to fulfill the requirements accuracy, reliability, and affordability. This time, there is a wide variety of approaches to reach the goal of higher accuracy without using expensive high-performance sensors. The algorithm presented in this paper utilizes image landmarks in combination with a low-cost Global Navigation Satellite System receiver, a detailed digital map, and vehicle odometry. In contrast to other approaches in the literature, which usually try to extract features from camera images and match those features with map information, the algorithm presented in this work moves this matching step from feature level to sensor data level. This means that map data is used to create an image of map features as the camera would sense it. The evaluation of this image prediction utilizes the coherency value, which is calculated from the structure tensor. By predicting a whole image, there is no more the need for a probabilistic description of extracted landmarks, i.e. it is also no decision about the existence of particular features necessary. Furthermore, any geometry including area landmarks can be used as landmarks. The paper describes a way to model area landmarks for efficient transformation to the predicted image. In the presented algorithm, the likelihood models used for the evaluation of the predicted coherency image are derived from real, manually labeled data. We show promising results of a test drive in an area with complex intersections. Those results are compared to ground truth data.