Essential Error Analysis of Self-Localization From Map Structure in Urban Environment

Abstract

Lane-level self-localization is a critical task in the field of autonomous driving. Map-based self-localization is commonly employed to achieve lane-level accuracy in urban settings. However, it is known that in certain locations, such as narrow roads and intersections, map-based self-localization poses challenges, leading to failures of autonomous driving tasks. Potential errors in specific locations can be estimated before driving by leveraging the geometrical structures of the environment and can then be utilized to enhance the safety of autonomous driving operations. In contrast to traditional self-localization error estimation methods, which often rely on statistical analyses or regression specific to a particular map format, this research focuses on identifying fundamental errors in self-localization within various map formats. The proposed method employs a general formulation of environmental representation, specifically normal distributions, and a closed-form uncertainty approximation of optimal solutions, enabling the identification of essential self-localization errors in an environment. The results obtained through this method are not only valuable for autonomous driving tasks but also contribute to discussions on the quality of specific map formats. Experiments highlight locations within an urban environment that are susceptible to significant errors in self-localization based on the proposed metric. Furthermore, a comparison of self-localization errors associated with specific map formats using the proposed metric reveals that essential errors in algorithms can be estimated. The discussion presented in this paper reveals the patterns of geometrical features in the urban environment that are likely to result in self-localization errors.

Keywords

• Autonomous driving
• self-localization
• high-definition maps
• uncertainty analysis