UC Berkeley

Technologies related to autonomous driving have been advancing rapidly for the past few years, and the community has already seen promising applications of self-driving vehicles in certain scenarios. However, urban driving environments, with their less structured roads, complicated traffic rules, and numerous visual occlusions, remain to be challenging. To assist the autonomous agents in such scenes, High Definition maps (HD maps) are proposed as an infrastructure for urban autonomous driving. A typical HD map usually has multiple layers of semantics including lane boundaries, drivable areas, and traffic rules. This prior knowledge of the scene is of great importance for downstream modules like perception, localization, prediction, and planning. Both in academia and in industry, nowadays, the HD map has become a fundamental module for autonomous driving.

With their loaded information, the HD maps, however, are not usually easy to obtain. Indeed, each module in the life cycle of an HD map imposes some challenges to researchers. The development of an HD map starts with the sensor setup on mobile mapping platforms and the collection of the mapping data. Typically, the mapping platform needs to be manually calibrated for the correct spatial-temporal relationship among sensors, and careful planning is required to prepare a mapping dataset. Then, the valuable data from these sensors are processed through HD map generation pipelines for future usage. Currently, the map construction requires intensive human labeling and post-processing, and few algorithms are robust enough to automatically map an arbitrary urban environment. Lastly, how to utilize a collection of HD maps is still a challenge for the mapping community. Due to limited onboard storage and computing capacity, loading a complete map would be infeasible, and end-users need an efficient submap query strategy in real-time.

Acknowledging the essential role of the HD map in urban autonomous driving and the difficulties in development, this dissertation discusses multiple perspectives in the complete life cycle of an HD map.

To begin with, this dissertation first discusses the mapping platform and the dataset for mapping applications in Part I. Chapter 2 focuses on the geometrical calibration and synchronization of the sensor suite on a mobile mapping platform. In this section, the complementary LIDAR-camera configuration is discussed, and a semantic-based optimization algorithm is proposed to estimate both the geometric and the temporal relationship between these two sensor modalities. In Chapter 3, an exemplar mapping platform and an urban dataset are introduced. The design of the mapping vehicle considers complicated urban scenarios, and the dataset includes some of the most challenging city driving scenes. The dataset is open to the public to encourage research in the mapping field. With the mapping platform configured, the next question in the life cycle of an HD map is the routing problem. With more than one mapping vehicle, how to efficiently route a mapping fleet is discussed in Chapter 4. Here, a Model Predictive Control-based algorithm is proposed to accommodate traffic conditions and map updating problems.

Part II focuses on the algorithms related to the automatic generation of the HD map. Chapter 5 introduces a particle filter-based algorithm to efficiently explore the lanes in complicated urban situations. The algorithm specifically solves the merging, forking, and irregular lane cases on city roads. Chapter 6 moves more towards the intersections and discusses potential solutions with a multi-sensor setup. Here, the mapping problem is treated as the semantic segmentation in the Bird's Eye View frame. Network design comparisons are also provided to demonstrate a preferred strategy in cross-domain fusion tasks. Chapter 7 studies the potential of multitask learning for both static and dynamic objects on the road to exploit information in limited data. Built upon a single backbone, the proposed method compresses six tasks into one neural network, and the evaluation shows that the performance is comparable with single-task models.

In Part III, the management and the deployment of the HD map are discussed. Chapter 8 introduces a tile-based map management system to query and combine smaller HD maps for real-time application. The proposed framework leverages an RB tree data structure and uses a submap queue during vehicle operation to store only useful maps onboard.

This dissertation concludes with a summary of breakthroughs in the life cycle of the HD map development process and comments on the future directions of HD map-related research.