Text Detection and Recognition on Traffic Panels From Street-Level Imagery Using Visual Appearance

Gonzalez, Alvaro; Bergasa, Luis M.; Yebes, J. Javier · 2014 · Crossref

DOI: 10.1109/tits.2013.2277662

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 challenge of automatically detecting and recognizing text on traffic panels from street-level imagery, a task complicated by the high variability in panel appearance, viewpoint deviations, and environmental conditions. While traffic sign recognition is well-studied, informative traffic panels remain difficult to process due to diverse content and lack of standardized databases. The authors propose a method to create automatic inventories of traffic panels for road maintenance and driver assistance systems, specifically targeting the Spanish road network using Google Street View images. The methodology employs a two-stage pipeline. First, traffic panel detection is performed using a Bag of Visual Words (BOVW) approach. To handle the visual variability, the system applies color segmentation masks to isolate blue and white regions, which are characteristic of Spanish traffic panels. Blue regions are segmented using a combination of three methods: a red-channel threshold to exclude the sky, hue-based filtering, and Otsu’s method on the difference between red and blue channels. White regions are detected using maximally stable extremal regions. Local descriptors are extracted from these masked regions using the Harris-Laplace detector and classified using either Support Vector Machines (SVM) or Naïve Bayes classifiers. Second, once a panel is detected, a text detection and recognition algorithm is applied. This stage includes a modified character recognizer capable of identifying letters, digits, and common traffic symbols. To improve accuracy, the system utilizes a language model based on a dynamic dictionary constrained by reverse geocoding, which limits potential word candidates to places relevant to the panel’s geographic location. The study was conducted on a dataset of 16,277 images from the Spanish road network, split into training and testing subsets. The training set included 5,514 images with 1,047 positive samples of 509 different panels, while the test set comprised 10,763 images. The authors evaluated various feature descriptors, including SIFT, C-SIFT, Hue-SIFT, RGB-SIFT, and color histograms, to determine the most effective representation for the BOVW model. The experimental results demonstrated the efficiency of the proposed visual appearance categorization method, showing that it could generalize to different scenarios without retraining. The integration of color segmentation with BOVW proved effective in distinguishing traffic panels from other road elements like advertisements or truck bodies. The significance of this work lies in its application to Intelligent Transportation Systems (ITS). By enabling the automatic creation of up-to-date traffic panel inventories, the method supports rapid evaluation of road signage conditions, such as visibility and deterioration. Furthermore, the system facilitates driver assistance and autonomous vehicle navigation by providing reliable text interpretation in areas where GPS coverage may be insufficient. The paper establishes that visual appearance-based methods, combined with geographic context, offer a robust solution for traffic panel recognition, overcoming the limitations of previous approaches that relied heavily on geometric features or lacked generalization capabilities.

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