In autonomous driving, achieving self-localization at the lane level is critical for various tasks, such as control and scene understanding. Particularly in urban areas, map-based methods have become the standard in recent years due to their stability. However, there are locations where map-based self-localization potentially performs poorly. Therefore, it is important to identify such locations in advance for a safe autonomous driving system. Existing methods for the prior evaluation of self-localization rely on prior experiments conducted in various settings. This paper proposes an evaluation metric focusing on an autocorrelation function of map structures, which is more intrinsic than existing evaluation methods. Specifically, we focus on the condition number of the autocorrelation function's shape to model the difficulty of identifying the true pose in the function. An experiment conducted in real urban environments revealed the presence of common areas where certain self-localization algorithms tend to fail. Furthermore, the experiment demonstrated a statistically significant difference in the proposed criterion between locations with high and low errors in self-localization, suggesting that criterion values can be used to identify areas with higher errors in self-localization.
Self-Localization Error Evaluation from Map Structure Based on Autocorrelation Values
2023-09-24
1048745 byte
Conference paper
Electronic Resource
English