Uncertainty-Aware Vision-based Risk Object Identification
via Conformal Risk Tube Prediction

We study object importance-based visual risk object identification (Visual-ROI), a key capability for detecting hazards in intelligent driving systems. Existing approaches are deterministic and ignore uncertainty, which can compromise safety. For example, using a fixed decision threshold in ambiguous scenarios would cause too-early or too-late detection of risks, and predictions that flicker between risky and non-risky states over time. These issues worsen under diverse contexts with multiple interacting risks, perturbing where and when risks occur. However, current vision methods lack a way to capture uncertainty jointly over space and time, limiting their ability to dynamically reflect changes in scene complexity.

We propose Risk Tube Prediction, a unified formulation that models spatiotemporal uncertainty in risk. We further introduce a new conformal prediction framework to provide coverage guarantees for the true risks and yield calibrated risk scores and uncertainty estimates. Specifically, we employ risk-category–aware calibrators that consider distinct characteristics to reduce confused calibration and localize risks more precisely in space and time. To evaluate, we present a new dataset and metrics probing diverse scenario configurations with multi-risk coupling effects. We systematically conduct experiments of factors that influence uncertainty estimation including variations in scenario configuration, per risk category analysis, and the propagation of perception errors. Our method delivers substantial improvements over prior approaches, enhancing both the robustness of Visual-ROI performance and downstream outcomes, such as reducing nuisance braking alerts.