Enhancing the Safety Horizon: Robust Uncertainty Quantification for Monocular Autonomous Navigation

Abstract

Autonomous agents in unstructured environments require accurate perception and reliable uncertainty quantification (UQ). Deep neural networks can model complex visual patterns but often become overconfident under distribution shift. This study presents an uncertainty-aware monocular navigability framework that combines semantic entropy from DeepLabV3, geometric uncertainty from Mi-DaS with Test-Time Augmentation (TTA), and Dempster-Shafer Theory (DST) fusion. Unlike prior under-specified fusion descriptions, this paper provides explicit mass construction, normalization, conflict handling, and safety-thresholding rules to improve reproducibility. On KITTI open-loop evaluation, depth uncertainty remains effective in the near field (ROC-AUC & 0.60 up to ˜17 m), while filtering high-uncertainty pixels in the 0-15 m braking-zone proxy reduces mean relative error by 21.5%. Semantic entropy achieves ROC-AUC 0.9991 for misclassification detection (ECE 0.1757). Runtime overhead and quantitative ablations against simpler predictors are also reported. Results support improved safety-oriented perception; closed-loop planning and collision-rate validation are left for future work.