Real-time situational awareness and collision risk evaluation for autonomous ships under sensor uncertainty

This study proposes an uncertainty-aware collision-risk evaluation framework for the safe navigation of autonomous ships. Target-ship information obtained from the Automatic Identification System (AIS) and radar is characterized by discrete update intervals and measurement uncertainty, which can degrade situational awareness and collision-avoidance decision-making. To address this problem, an Unscented Kalman Filter (UKF) was used to fuse AIS and radar data and estimate target-ship states and associated uncertainties in real time. The resulting state-estimation covariance was incorporated into the Quaternion Ship Domain (QSD) to construct a probabilistic ship domain that reflects target-position uncertainty. Based on this domain, Probabilistic Collision Exposure (PCE) was defined as a velocity-space collision-exposure metric representing the proportion of collision-causing velocity vectors within the own ship's reachable velocity set, while Expected Collision Energy (ECE) was formulated to incorporate the potential collision consequence. The framework was evaluated through MMG-based simulations involving head-on, starboard crossing, nonlinear maneuvering, and three-target multiple-encounter scenarios under delayed and noisy sensing conditions. In the nonlinear maneuvering scenario, the target-position error remained within approximately 5 m, with an average error of approximately 3 m. Compared with the conventional Collision Risk Index (CRI), PCE showed substantially lower residual risk after the closest point of approach (CPA) in the representative head-on and starboard crossing scenarios. The post-CPA residual area-under-the-curve (AUC) ratios were 2.21% and 0.59% for PCE in the head-on and starboard crossing scenarios, respectively, compared with 32.80% and 40.48% for CRI. The three-target scenario further demonstrated target-specific collision-exposure evaluation and simultaneous multi-target avoidance, while the implemented framework maintained its nominal 10-Hz system update cycle across the evaluated single-target and multi-target simulations.