A novel traffic sign recognition approach for open scenarios

Weipeng, CAO; Yuhao, WU; Dachuan, LI; Zhong, MING; Zhenru, CHEN; Xuan, YE · 2023 · DOAJ

DOI: 10.3724/SP.J.1249.2023.03258

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 instability and security risks of traditional deep learning-based traffic sign recognition (TSR) systems in open-world scenarios. Conventional models rely on fully data-driven, end-to-end training under independent and identically distributed assumptions, lacking the ability to reason about unseen categories or explain decisions. This limitation is critical in real-world applications where traffic signs may be dynamically updated or where certain classes are missing from training data due to collection difficulties. To mitigate these issues, the authors propose a novel TSR framework that integrates domain knowledge from national traffic sign design standards with zero-shot learning (ZSL) mechanisms. The methodology involves constructing a semantic dataset by abstracting general attributes of traffic signs—specifically shape, color, icon, character, and function—from design standards. These attributes are encoded into semantic vectors and injected into the training process as prior knowledge. The proposed framework utilizes a backbone network (ResNet-101) to extract visual features, which are then mapped to semantic attribute vectors via a semantic auto-encoder (SAE) or an SAE with L2-norm constraint (SAE-L2). During inference, the model predicts the semantic vector of an input image and matches it against preset category semantic vectors using similarity metrics, such as cosine similarity. This approach allows the model to recognize signs not seen during training and provides interpretable decision paths based on attribute mismatches. Experiments were conducted on the Chinese Traffic Sign Database (CTSDB) and the German Traffic Sign Recognition Benchmark (GTSRB). The study evaluated performance under both traditional ZSL settings (testing only on unseen classes) and generalized ZSL settings (testing on both seen and unseen classes). Results demonstrated that the proposed framework significantly outperformed random prediction and traditional deep learning models in recognizing unseen traffic signs. Specifically, in traditional ZSL settings, the accuracy for unseen signs improved by at least 29.96% on CTSDB and 24.25% on GTSRB compared to random prediction. The SAE-L2 model consistently showed better generalization capabilities than the standard SAE model. However, the authors note that while the framework proves feasible and effective, the absolute accuracy rates have not yet reached levels suitable for practical deployment, indicating a need for further algorithmic improvements. The significance of this work lies in its "knowledge + data" driven approach, which enhances the robustness and interpretability of TSR systems in open scenarios. By leveraging design standards as domain knowledge, the method addresses the challenges of class imbalance and missing categories, offering a scalable solution for dynamic traffic environments. The study validates the feasibility of using ZSL mechanisms to bridge the gap between controlled training environments and unpredictable real-world applications, providing a foundation for future research in trustworthy autonomous driving systems.

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