Reference-Free Positional Accuracy Evaluation for Crowdsourced HD Map

High-efinition (HD) maps are essential for autonomous driving (AD). While crowdsourcing is an effective and low-cost method for creating and updating HD maps, limitations in on-board sensors and environmental factors can result in severe map quality issues. Consequently, quality assessment is crucial to ensure the reliability of maps used in AD systems. In this study, we propose a framework for evaluating the positional accuracy of landmarks locations in crowdsourced HD maps without the need for reference data. The framework consists of two main components: in the first part, we introduce a semi-supervised clustering model to assess GNSS signal confidence, estimate the signal strength of different road sections, and support scene detection. In the second part, the map quality assessment problem is decomposed into several one-dimensional quality classifiers based on empirical rules. We introduced ground truth to validate the results of our algorithm. The results indicate that our predictive results show improvements over traditional methods across various scenarios.