Extraction, localization, and fusion of collective vehicle data

Abstract

Maps representing the detailed features of the road network are becoming more and more important for self-driving vehicles and next generation driver assistance systems. The mapping of the road network, by specially equipped vehicles of the well-known map providers, leads to usually quarterly map updates, which might result in problems encountered by self-driving vehicles in the case that the road information is outdated. Furthermore, the provided maps could lack details, such as precise landmark geometries or data known to exhibit a fast temporal decay rate, which might be, nevertheless, highly relevant, such as friction data.

As an alternative, extensive amounts of information about the road network can be acquired by common vehicles, which are, nowadays, commonly equipped with manifold types of sensors. Subsequently, this type of gathered data is referred to as CVD (Collective Vehicle Data). The process of map creation requires, at first, the extraction of relevant sensor data at the vehicle-side and its accurate localization. Unfortunately, sensor data is typically affected by measurement uncertainties and errors. A minimization of both can be achieved by means of an appropriate sensor data fusion.

This work aims for a holistic view of a three-staged pipeline, consisting of the extraction, localization, and fusion of CVD, intended for the derivation of large-scale, high-precision, real-time maps from collective sensor measurements acquired by a common vehicle fleet. The vehicle fleet is assumed to be solely equipped with commercially viable sensors. Concerning the processing at the back-end-side, general approaches that are applicable in a straightforward manner to new types of sensor data are strictly favored. For this purpose, a novel distinction of CVD into areal, point-shaped landmark, and complex landmark data is introduced. This way, the similarities between different types of environmental attributes are exploited in an overall highly beneficial manner; and the proposed algorithms can be adapted to new types of data that appertain to these categories by appropriately adjusting their parameterizations. To achieve the above mentioned goals, both novel approaches, where the research lacks established ways, and relevant extensions/adaptations of existing ones are suggested to fulfill the very specific, automotive requirements.

All in all, the thesis condenses a broad and manifold research concerning the deduction of large-scale and high-precision map data grounded on preprocessed sensor measurements that have been acquired by common vehicles, the so-called CVD. The focus is put on the utilization of commercially viable sensors. Additionally, besides its broad perspective, this thesis also emphasizes highly relevant details, such as the efficient, adaptive temporal weighting of sensor data at the back-end-side and the template-based hierarchical data storage. A complete pipeline, consisting of the extraction, localization, and fusion of CVD, is presented and evaluated, as each component is known to have a direct impact on the quality of the deduced map data. Approaches to the fusion of areal and point-shaped/complex landmark data are either invented from scratch or significantly enhanced according to the state of the art, always bearing in mind the highly specific needs of the automotive context.