Ontology-based map modelling and processing for autonomous vehicles

Abstract

Autonomous Driving (AD) systems use digital maps as a virtual sensor to anticipate the road ahead and make decisions. The evolution of digital maps has posed a range of challenges to the current mapping ecosystem. First, different standards and formats make maps lack interoperability. Second, the regular update and increasing size of the maps data make it largely impossible to store and process a complete detailed map in a navigation system of a car. Finally, errors introduced in any step of map creation may cause unavailability of AD functions. Currently, there are separate software components trying to solve each challenge, however, the integration and maintenance of these components is a difficult task due to the challenges of the maps' evolution. This thesis investigates an ontology-based approach with the goal of shifting from map data-oriented functional design to knowledge-centered ontology design. Thus, we contribute a knowledge-spatial architecture with an embedded quality assurance mechanism to achieve efficient dynamic map provision with quality assurance, providing flexible query answering.

Our first contribution is two levels of ontological abstraction to solve the map data integration problem. The developed low-level ontologies represent the specific map data formats. A single high-level ontology is designed with the prerequisites and requirements of a self-driving vehicle in mind resulting in a light(er) weight generic map ontology. Mapping rules are designed to unify low-level map ontologies to the generic map ontology. As our second contribution, we develop and implement an efficient dynamic map update strategy to provide continuous road knowledge ahead. To achieve efficient map updates, we design a spatial-sliding window on top of the light(er) weight generic map ontology and process map data streams via reasoning based on a pre-fetching mechanism. To ensure map data quality and preventing AD mode degradation, we conduct our third contribution which is the design of a workflow to detect and fix the map data violation including semantic enrichment, violation detection, and violation handling. To facility violation detection, we develop a Map Quality Violation Ontology and a set of constraint rules. Violation handling is realized based on a real-world map data error.

To evaluate the proposed methods, we rely on empirical evaluations as well as on the development of concrete use cases. The attained results provide evidence that an ontology-based approach enables effective map integration and processing with ensured data quality. This allows engineers to focus more on developing AD functions on the knowledge level rather than on data processing and integration. The proposed two-level ontology development methodology shades the light for ontology practitioners to build ontologies in situations where the data is dynamic, and the computation resources are limited.