Human-inspired risk level annotation for autonomous driving in open-pit mine environments

Ensuring safe and efficient autonomous driving in open-pit mines remains a significant challenge due to the complexity and risk of these environments. Although current risk assessment approaches can quantify hazards, they often fail to enable autonomous driving system (ADS) to accurately perceive and interpret scene risks as human do, limiting their interpretability and practical applicability. To address this, we propose a human-inspired risk assessment framework for ADS in open-pit mines, supported by a novel annotated dataset and DS-based label fusion. The Dynamic Bayesian Network (DBN)-based risk assessment framework provides a comprehensive evaluation of potential risks and incorporates an interaction risk coefficient to account for collision severity between vehicles of different masses. A crowd-sourced, expert-annotated dataset is constructed, with inter-rater reliability and multi-source label fusion ensured via Fleiss’ kappa and the Dawid–Skene model. Experiments show that the framework improves interpretability, promotes safer driving behaviors, and facilitates practical deployment of ADS in open-pit mining environments.