More about HKUST
Automatic Generation of Vector Maps by Imitation Learning for Autonomous Driving
The Hong Kong University of Science and Technology Department of Computer Science and Engineering PhD Thesis Defence Title: "Automatic Generation of Vector Maps by Imitation Learning for Autonomous Driving" By Mr. Zhenhua XU Abstract The vector map of road elements is critical for autonomous driving. It contains the connectivity and topology information of the road for multiple downstream tasks of the autonomous vehicles, such as prediction, planning and motion control. To assure the normal and safe operations of autonomous vehicles, vector maps, including the SD (standard-definition) map and the HD (high-definition) map need to be pre-built before the deployment of autonomous vehicles in the target region. However, manually annotating the vector map is labor-intensive, expensive and time-consuming, which severely restricts the fast application and deployment of the autonomous vehicle. Therefore, automatic approaches to effectively and efficiently generate the vector map of the target region are needed. The source data may come from high-resolution aerial images or vehicle-mounted sensors, such as cameras or LiDARs. Aerial images can directly provide high-quality BEV (bird's-eye view) images around the vehicle, while additional steps for perspective transformation are required for vehicle-mounted sensors. Previous works on our task could be classified into three categories: segmentation-based approaches, two-stage graph-based approaches and decision-making graph-based approaches. Compared with segmentation-based approaches and two-stage graph-based approaches, the decision-making graph-based approaches are more powerful and stable. Therefore, this thesis mainly focuses on decision-making graph-based approaches. These approaches are trained by imitation learning. In this thesis, we will discuss three decision-making graph-based systems for vector map generation with details. First, we propose to detect the graph of road curbs from small aerial image patches with a decision-making network trained by imitation learning. Road curbs possess simple topology without complicated intersections or overlapping. Then, we manage to detect the graph structure of the road network, whose topology is much more complicated than road curbs and road boundaries, such as intersections of arbitrary numbers of roads and overpasses. A decision-making system based on the transformer is proposed to generate the vector map vertex by vertex. The proposed system can handle road networks with any topological changes, which presents state-of-the-art performance compared with previous works. Finally, we adapt and further improve our approaches to generate the vector map of road lane centerlines with data collected by vehicle-mounted sensors. We first convert the data from perspective views to BEV by perspective transformation. And then generate the global consistent vector map by the proposed iterative vectorization approach. The proposed system achieves state-of-the-art performance in evaluation experiments. In the end, we analyze the remaining problems that cannot be well solved by existing approaches as potential future works, such as the map update problem, adapting more advanced decision-making algorithms to our task, and graph merge. Date: Friday, 6 January 2023 Time: 2:00pm - 4:00pm Venue: Room 3494 lifts 25/26 Chairperson: Prof. Zhihong GUO (CHEM) Committee Members: Prof. Ming LIU (Supervisor) Prof. Huamin QU (Supervisor) Prof. Qifeng CHEN Prof. Qiong LUO Prof. Yansong YANG (ECE) Prof. Yong WANG (Harbin Institute of Technology) **** ALL are Welcome ****