Out-of-Distribution Image Detection for Robust Object Detection in UAV Platforms

Abstract

[Abstract]: Object detection systems deployed on unmanned aerial vehicles (UAVs) operate in dynamic and safety-critical environments where visual perception is frequently degraded by motion, weather, illumination changes, and sensor artifacts. These distributional shifts can silently compromise detection reliability, increasing the risk of perception failures in autonomous flight. This thesis proposes a structured framework for evaluating robustness in UAV-based object detection and for predicting perception failures through out-of-distribution (OOD) detection. First, the robustness of a YOLOX-based aerial object detector is systematically evaluated under a taxonomy of UAV relevant synthetic corruptions, identifying near-OOD conditions that induce significant detection degradation while preserving scene semantics. Second, a lightweight background classification model is trained to learn environmental context representations, enabling the modeling of both appearance-based near-OOD and semantically novel far-OOD scenarios. Multiple state-ofthe-art OOD detection methods, including ODIN, Mahalanobis distance scoring, Gram matrix deviation, Generalized ODIN, and energy-based approaches are then benchmarked under aerial distribution shifts. Experimental results demonstrate that Energy-Bounded Learning achieves the strongest and most consistent separation between in-distribution and OOD samples across both covariate and semantic shifts. Finally, OOD scores are calibrated into a continuous probabilistic estimate of epistemic risk, enabling interpretable, real-time reliability assessment in UAV video streams. By reframing OOD detection as a proxy for perceptual failure prediction, this work contributes a practical and computationally efficient mechanism for enhancing robustness, uncertainty awareness, and operational safety in autonomous aerial perception systems.