Autonomous Vehicle Localization without Prior High-Definition Map

Abstract

Accurate localization by which vehicles can arrive at their destination while accurately following a given route is one of the most important factors for autonomous driving. In recent years, numerous studies have been conducted to achieve accurate localization using high-definition (HD) maps. Based on the HD map information (e.g. spatial data, lane, and traffic sign), autonomous vehicles can localize themselves by matching the surrounding spatial information obtained from onboard sensors to the HD maps. However, generating HD maps is a time-consuming and costly task. This study introduces a time-saving, effective, and accurate localization method inspired by humans, using only onboard sensors and publicly available two-dimensional (2D) map information. Similar to the multi-level localization process performed by humans, the proposed method interprets and matches the surrounding spatial data to the publicly available 2D maps using deep-learning-based place recognition and simultaneous localization and mapping (SLAM), thereby enabling autonomous vehicles to localize even without prior HD maps. Through the proposed method, our framework enables autonomous vehicles to perform maximally decimeter-level accurate localization without using HD maps. Evaluation of the proposed method using various datasets and publicly available map sources demonstrates that accurate global localization can be achieved without prior HD maps.