Explore a groundbreaking approach to near-real-time crowd-sourced 3D map collection for autonomous vehicles in this 20-minute conference talk from NSDI '20. Delve into CarMap, an innovative system that addresses the challenge of creating compact, up-to-date 3D maps for vehicle localization. Learn about its lean representation of feature maps, novel object filtering techniques, and position-based feature matching to enhance localization robustness. Discover how CarMap's stitching algorithm combines map segments from multiple vehicles for unmapped regions and efficiently updates existing map areas. Gain insights into CarMap's performance, including its ability to update maps in less than a second, reduce map sizes by 75× compared to competing strategies, and improve localization accuracy in challenging scenarios.
Overview
Syllabus
NSDI '20 - CarMap: Fast 3D Feature Map Updates for Automobiles
Taught by
USENIX