Examenspresentation av Yubo Tian och Runjia Qian

Examinator: Lars Hammarstrand 

Abstract:
The creation and updating of high-definition (HD) maps have become vital for advanced applications like autonomous vehicles, yet traditional map-making techniques face inherent limitations in cost, efficiency, and scalability. Given the same challenge and goal, this thesis explores two distinct approaches, the optimization-based and the learning-based methods, both leveraging the wealth of crowdsourced data to enhance HD map construction and update processes.

The optimization-based method utilizes a three-stage process involving the handling of raw data, the alignment and optimization of maps from multiple drives including lane marker and traffic sign alignment, and the regular generation and updating of maps. Alternatively, the learning-based method employs a two-stage network that uses a vision transformer as a crowdsourced data aggregator and a detection transformer for decoding individual road elements, linking neighboring frames using pseudo labels. In conclusion, both methodologies present robust, cost-effective, and time-efficient solutions for HD map creation and updates, showcasing significant advancements in mapping techniques through the combined power of optimization and machine learning.

Welcome!

Yubo, Runjia and Lars