Robustness and Noise Evaluation for Real-World Object Detection: Defending Against Sensor Degradation

Object detection models deployed in industrial and autonomous edge systems frequently encounter severe sensor degradation, notably high-frequency Gaussian noise introduced by low-light camera sensors, thermal thermal drift, and analog-to-digital converter noise. This study establishes a rigorous empirical evaluation framework for quantifying the degradation of YOLOv8n under controlled synthetic Gaussian sensor noise across six standard deviation regimes (sigma in [0.0, 0.5]). Across repeated multi-seed simulations on the COCO128 benchmark, we observe a non-linear performance collapse: clean baseline mAP@50 degrades from 0.6073 +- 0.0019 to 0.2188 +- 0.0056 at sigma=0.5 (a 64.0% relative drop), accompanied by a parallel collapse in object recall from 0.6356 to 0.2306. We provide 95% bootstrap confidence intervals across 5 independent simulation seeds, identify critical failure boundaries for edge deployment, and release full reproducibility scripts and raw datasets.