Uncertainty Quantification in Learning-Based Perception for Safety-Critical Autonomous Systems: Methods and Applications

Abstract

Learning-based perception systems using deep neural networks have achieved remarkable performance in autonomous systems, yet their deployment in safety-critical applications requires rigorous uncertainty quantification to identify when predictions may be unreliable and trigger appropriate safety responses. This paper comprehensively examines methods for uncertainty quantification in perception systems and their applications across unmanned system domains. We systematically distinguish aleatoric uncertainty arising from inherent sensor noise and environmental variability, from epistemic uncertainty reflecting model limitations and insufficient training data. The review analyzes uncertainty quantification approaches including Bayesian neural networks using weight distributions rather than point estimates, Monte Carlo dropout approximating Bayesian inference through stochastic forward passes, ensemble methods aggregating predictions from multiple models, and evidential deep learning directly predicting uncertainty through evidential distributions. Particular emphasis is placed on computational efficiency for real-time deployment, examining approximation techniques and hardware acceleration strategies. We evaluate calibration quality—the alignment between predicted confidence and actual accuracy—using reliability diagrams and expected calibration error metrics. Application-specific implementations are analyzed for object detection with uncertainty-aware bounding boxes, semantic segmentation with per-pixel confidence estimates, and depth estimation with uncertainty maps. The paper examines integration of uncertainty information into downstream decision-making including risk-aware planning that avoids high-uncertainty regions, active learning selecting informative samples for labeling, and anomaly detection flagging out-of-distribution inputs. Validation methodologies are discussed including testing on corrupted data, adversarial examples, and domain shift scenarios. The review identifies research gaps in uncertainty quantification for temporal models, multi-modal fusion, and providing formal safety guarantees based on uncertainty estimates.