Benchmarking real-world physical tampering on traffic signs and evaluating its impact on recognition models

Abstract

Traffic Sign Recognition (TSR) is a core perception component in advanced driver-assistance systems and autonomous vehicles, where even infrequent recognition failures can lead to severe safety consequences. Although modern TSR models demonstrate high accuracy on widely used benchmarks, their robustness under real-world physical tampering of traffic signs remains insufficiently studied and poorly standardized. In practical environments, traffic signs are frequently affected by stickers, graffiti, dirt, fading, abrasion, partial occlusion, deformation, and reflective surface alterations, all of which can significantly distort visual cues used by recognition models. This paper proposes a comprehensive benchmarking framework for evaluating TSR performance under real-world physical tampering conditions. The framework introduces a structured taxonomy of tampering types, severity levels, and an annotation protocol applicable to both still images and short driving sequences. In addition to conventional detection and classification accuracy, the benchmark emphasizes temporal stability metrics to capture recognition flicker and persistent misclassification across consecutive frames. An experimental evaluation design is presented to compare baseline TSR models with tamper-aware robustness strategies. The proposed benchmark aims to bridge the gap between laboratory evaluation and real-world deployment, enabling more reliable and reproducible assessment of TSR systems under realistic operating conditions..