Abstract Maritime collision risk assessment is essential for ensuring navigational safety under increasingly congested traffic conditions. In such contexts, autonomous vessels must evaluate risks efficiently and respond in real time, which remains challenging when using conventional methods. The widely used Dempster-Shafer (D-S) model, although effective in theory, suffers from high computational complexity and limited scalability when applied to multi-vessel encounters involving both Maritime Autonomous Surface Ships (MASS) and conventional vessels. To address this issue, a machine learning-based framework is proposed, in which an Extreme Gradient Boosting (XGBoost) model replaces the theoretical D-S model for Collision Risk Indicator (CRI) estimation. Automatic Identification System (AIS) data are used to construct realistic simulation scenarios, and the predicted CRI is continuously evaluated and integrated into an automatic collision avoidance algorithm. The proposed model achieves an R² of 95.36% during training and 90.21% in real-world simulation testing. In high-traffic scenarios involving more than 20 vessels, it demonstrates significantly faster processing speed than the D-S model. Safety analysis further shows that integrating CRI with the Velocity Obstacle (VO) algorithm reduces collision risk by 33.0% in autonomous-autonomous vessel encounters and 28.7% in autonomous-conventional vessel encounters. These results indicate that the proposed method supports efficient and scalable real-time collision risk management for autonomous maritime navigation.
Li et al. (Wed,) studied this question.