دانلود کتاب یادگیری از روی نقشه‌ها: تولید مقیاس‌پذیر «حقیقت مبنا» (Ground Truth) در رانندگی خودکار

A reliable and robust map perception is a key enabler for scalable autonomous driving by supporting multiple essential functions: Detecting map elements from on-board sensors enables localization within a highdefinition (HD) map, verification that a map is up-to-date, and on-the-fly perception of map information towards map-less driving paradigms. Traditional methods for training perception systems rely heavily on the costly, time-consuming, and often sensor-specific process of manually annotating sensor data. In contrast, this work presents a scalable approach for the generation of training data for machine learning models by leveraging HD maps as a supervision source across different sensor modalities and perception tasks. The first part of the thesis details the specification, recording and processing of a multi-drive dataset that serves as a foundation for learning perception models directly from HD maps. Complementary mapping approaches are applied to create a comprehensive set of map layers. The contained map elements differ in geometric representation, abstraction level, and functional role and offer a diverse set of training targets for various perception tasks. Based on this foundation, the second part of the thesis trains and evaluates multiple perception models across different sensor modalities and learning representations. Towards this goal, a bird’s-eye view (BEV) perception model for online HD map construction from surround-view cameras is trained, while analyzing and optimizing the training regime across multiple dimensions. Complementary to the BEV representation, multiple perspective view map perception models are trained for different sensor modalities and tasks. One line of experiments generates pixel-accurate annotations for front-view cameras, where trained models robustly predict diverse map element classes, including fine-grained map elements such as dashed road markings, traffic lights, and traffic signs. The effectiveness of the proposed approach is also demonstrated in a fully self-supervised cross-modal domain adaptation setting, transferring knowledge from a richly annotated image dataset to the Light Detection and Ranging (LiDAR) domain via automatically generated HD maps.