Out-of-Distribution Detection Failures in Autonomous Perception Systems: A Failure Mode Analysis

Abstract

Autonomous perception systems rely on out-of-distribution (OOD) detection to prevent unreliable inputs from entering downstream decision processes. However, detector failures are rarely uniform across deployment conditions. This study analyzes OOD detection failures in autonomous perception through a scenario-stratified failure mode analysis covering night scenes, heavy fog, unseen objects, sensor artifacts, and foreign road layouts. The results show that unseen objects achieved the strongest detector performance, with AUROC of 0.921, AUPR of 0.884, and FPR95 of 16.4%, indicating that semantic novelty was comparatively easier to separate. Sensor artifacts produced the weakest performance, with AUROC of 0.756, AUPR of 0.701, and FPR95 of 34.7%, followed by heavy fog with AUROC of 0.804 and FPR95 of 29.1%. Failure distribution analysis showed that false acceptance dominated safety-critical scenarios, reaching 31.4% in sensor artifacts, 27.8% in heavy fog, and 24.6% in foreign road layouts. Calibration analysis further revealed severe confidence drift, with Expected Calibration Error increasing from 0.038 in in-distribution samples to 0.167 under sensor artifacts and 0.142 under heavy fog. Risk prioritization identified false acceptance and sensor-noise failure as the highest-priority modes, each reaching a risk score of 80. Downstream analysis showed that accepted sensor artifacts reduced object accuracy to 58.9% and depth consistency to 55.7%, while foreign road layouts reduced lane stability to 57.6%. These findings demonstrate that OOD detection must be evaluated as an embedded safety mechanism rather than an isolated classifier. The study contributes a failure-oriented framework linking detector metrics, calibration behavior, scenario-specific brittleness, downstream perception degradation, and operational risk prioritization.