Perception uncertainty as an early indicator of image sensor aging

Camera imaging systems inevitably undergo aging during operation. Degradation can occur both in optical and in electrical components of visual systems. In this work we focus on cumulative sensor aging in the form of hot-pixel formation at the sensor level over the device lifetime. The main contribution is the integration of an empirical hot-pixel aging law with perception-level uncertainty analysis, enabling sensor aging to be linked to detector output instability. To evaluate perception stability, we apply Monte Carlo simulations across real automotive images and four object detection architectures. To validate the framework beyond simulation, we additionally evaluate it on the large-scale nuScenes dataset, which provides multi-camera data with ground-truth annotations. This enables evaluation across multiple camera views and modern CMOS imaging systems, complementing the controlled raw CCD-based analysis. The results show that both confidence and localization variance increase with simulated sensor age, while mean detection accuracy remains largely unchanged in early stages of degradation. In the large-scale experiments, confidence variance grows by approximately 6 × – 7 × between 5 and 20 simulated years, while the corresponding mean recall at a common confidence threshold remains stable within ± 1 % . These results confirm that uncertainty-based metrics provide an early indicator of sensor degradation before conventional accuracy metrics degrade, and demonstrate that the observed behavior generalizes across different sensor technologies, camera viewpoints, and datasets. This study is conducted under simulation of sensor degradation and serves as a proof-of-concept analysis of perception instability under physically motivated aging processes, with applicability to the evaluation of camera-based perception systems in autonomous driving, robotics, and related domains.