| Summary: | Digital street maps with rich features are the foundation of many applications. However, creating and maintaining up-to-date digital maps often involve many labor-intensive tasks, making the mapping process time-consuming and expensive. This thesis explores automated techniques for enriching digital street maps from aerial imagery and GPS data.
Digital street maps consist of a collection of geometry structures such as a road graph and the semantics associated with the structures, such as the lane count and the speed limit of a road segment. This thesis first proposes two solutions, RoadRunner and Sat2Graph, to automatically extract road-level street maps from GPS trajectory data and aerial imagery, respectively. Road-level street maps serve as the base maps in digital street maps, providing the basic yet fundamental way-finding service to the map users. However, road-level street maps don't have lane structure information, which is essential for lane-to-lane navigation and autonomous vehicles. Therefore, this thesis proposes a mapping pipeline that extracts lane-level street maps from aerial imagery. Besides road structure extraction, this thesis proposes RoadTagger to infer road attributes such as the lane count and road type of road segments from aerial imagery. Finally, this thesis proposes a mapping solution to create high-resolution traffic accident risk maps that can enrich the semantics of existing digital maps and enable new applications such as safety-aware routing and precise insurance.
|
|---|