Title:	Creating HD maps for autonomous driving
DNr:	LiU-compute-2023-40
Project Type:	LiU Compute
Principal Investigator:	Yuxuan Xia <yuxuan.xia@liu.se>
Affiliation:	Linköpings universitet
Duration:	2023-11-26 – 2024-12-01
Classification:	10207
Keywords:

Abstract

Autonomous driving has the potential to revolutionize transportation by making it safer, more efficient, and environmentally friendly. In realizing this vision, one of the key challenges is the construction of high-definition (HD) maps, which are roadmaps with centimeter accuracy and high fidelity. HD maps enable self-driving vehicles to perform precise localization and make informed decisions about how to navigate through complex environments. However, the creation and maintenance of HD maps typically requires the use of expensive mapping systems with heavy manual modifications. An economically efficient and highly scalable solution to address this problem is to use crowdsourced mapping by leveraging massively available crowdsourcing devices. Crowdsourcing devices typically include low-cost cameras and global navigation satellite system sensors, installed on passenger vehicles that frequently travel on the road. Therefore, a large amount of road observation data can be easily obtained. In crowdsourced mapping, vehicles equipped with crowdsourcing devices automatically collect anonymous data as they drive. Two possible architectures for crowdsourced mapping are centralized fusion and decentralized fusion. In centralized crowdsourced mapping, the collected data undergoes preprocessing to extract pertinent information, which is subsequently transmitted to the cloud server for updating the global HD map. Through aggregation and alignment of data from individual drives along the same road, the road model is inferred, and the HD map is updated. In decentralized crowdsourced mapping, each vehicle first processes the data and updates the HD map locally, and then the updated local HD maps are incrementally fused in the cloud server to obtain the updated global HD map. Compared to a centralized architecture, a decentralized architecture is more cost-effective and scalable, but it can also be difficult to ensure accuracy. While crowdsourced mapping has received increasing attention in recent years, there remain many challenges to resolve. The main objectives of this project are to develop new models for distinct features in HD maps as well as novel algorithms for creating and maintaining HD maps for autonomous driving with crowdsourced data, advancing the state-of-the-art.