Reliable Traffic Sign Recognition System

Atif, Muhammad; Ceccarelli, Andrea; Bondavalli, Andrea · 2021 · Crossref

DOI: 10.5753/ladc.2021.18528

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical need for robust Traffic Sign Recognition (TSR) systems within Advanced Driving Assistance Systems (ADAS). While TSR is essential for driver assistance and autonomous driving, misclassification poses severe safety risks. The authors identify that existing TSR systems, which often rely on single-frame processing, are vulnerable to real-world failures such as camera defects (broken lenses, dead pixels), occlusions, motion blur, and physical alterations to signs. The research is motivated by the lack of comparative studies evaluating TSR performance under these ongoing failure conditions. The goal is to develop strategies that enhance system reliability against such adversarial and environmental challenges. The proposed methodology involves generating synthetic faulty datasets using Generative Adversarial Networks (GANs) and manual corruption techniques, such as inserting random patches, to simulate real-world failures. These corrupted images are derived from three benchmark datasets: the Dataset of Italian Traffic Signs (DITS), the German Traffic Signs Recognition Benchmark (GTSRB), and the BelgiumTSC dataset. The authors explore two primary strategies to improve robustness. First, they utilize data augmentation by including faulty images in the training phase to help classifiers learn to handle corrupted inputs. Second, they design a Failure Detector (FD), implemented as either a binary or multi-class classifier using traditional machine learning algorithms (KNN, SVM, Random Forest) and deep learning models (AlexNet, Inceptionv3, MobileNetv2). The FD interacts with the TSR component through three distinct architectures: pre-filtering images, enriching TSR feature sets with FD probabilities, or using a meta-level classifier to combine outputs from parallel FD and TSR systems. The study reports that a sliding window approach for multi-frame TSR, combined with a deep classifier and a meta-level classifier, achieved perfect classification accuracy across all three benchmark datasets. Additionally, the authors evaluated the robustness of deep classifiers against specific camera failures, identifying certain types of camera defects as particularly dangerous and requiring further attention. The experimental design aims to benchmark these approaches to determine the most effective method for mitigating failure impacts. The significance of this work lies in its comprehensive strategy for building reliable TSR systems capable of operating under imperfect environmental and hardware conditions. By integrating failure detection and data augmentation, the authors conjecture that TSR systems can achieve higher robustness and safety. The paper concludes that combining a Failure Detector with the TSR component offers a promising pathway for creating unified, reliable recognition systems, with ongoing experimental campaigns planned to further validate these conjectures.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• sign visibility legibility