We present an edge-AI monitoring system that reliably detects long-term performance degradation of object detectors. The core is a localisation-aware post-hoc calibration that aligns a detector’s confidence with IoU-based localisation accuracy between predicted and ground-truth boxes. Concretely, raw scores pipi are transformed via Platt Scaling into calibrated probabilities... which are consistent with localisation accuracy. To track drift, we apply temporal chunking and evaluate each chunk tt with a sliding step, computing the mean calibrated confidence... We then perform chained thresholding, updating the reference level by ... which makes the system robust to short-term noise yet sensitive to long-term drift. On UA-DETRAC with day–night domain shift and scene-wise splits, the proposed method improves localisation-aware calibration error (LaECE00) while achieving stable degradation detection, and it operates as a lightweight post-processing layer applicable across detectors and datasets.

Dongyoung Lee was born in Anyang, South Korea, in january 2000. He earned his bachelor's degree in Biomedical Engineering from Gachon University in February 2025. He is currently pursuing a master's degree in AI Imaging at the Graduate School of Advanced Imaging Science, Multimedia & Film, Chung-Ang University, Seoul, South Korea. His research interests include computer vision