A Real-Time Road Sign Detection Using Bilateral Chinese Transform

Belaroussi, Rachid; Tarel, Jean-Philippe · 2009 · Crossref

DOI: 10.1007/978-3-642-10520-3_111

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 real-time road sign detection in still images, aiming to overcome the limitations of existing methods such as the Radial Symmetry Transform (RST) and standard color segmentation. Previous approaches often suffer from high false positive rates, sensitivity to missing edge points, and a lack of generality, typically requiring specific shape detectors or strict color models. The authors propose the Bilateral Chinese Transform (BCT), a pairwise gradient-based symmetry transform capable of detecting both circular and polygonal road signs regardless of their pose, orientation, or contrast polarity (light-on-dark or dark-on-light). The BCT method operates on intensity images, specifically using a normalized red channel to mitigate illumination variations without relying on rigid color models. Unlike the standard Chinese Transform (CT), which assumes objects have a bright-on-dark contrast, the BCT modifies the phase weighting function to account for both convergent and divergent gradient pairs. This allows it to detect signs with varying contrast types. The algorithm uses a pairwise voting scheme where edge points vote for their midpoint based on radial symmetry and alignment conditions. It employs distance thresholds rather than discrete scales to define the spatial extent of detected objects, calculating the radius as an arithmetic average of voter distances. This approach makes the detector orientation-free and robust to in-plane and small out-of-plane rotations. Experimental results were evaluated on a database of 89 images (640x480 resolution) containing 92 road signs of various colors and shapes, excluding triangular signs. The BCT achieved a detection rate of 86% (79 out of 92 signs) with 25 false positives, processing each image in approximately 30 milliseconds. In comparison, the standard CT detected only 75% of signs with 24 false positives, while a color-segmentation-based approach (CT with Camshift) detected 75% of red signs but missed many non-red signs due to color model mismatches. The BCT demonstrated a higher Dice coefficient (82%) compared to the standard CT (77%), indicating superior segmentation accuracy. The study highlights that BCT is more general and precise than RST and GST, effectively handling diverse sign colors and shapes without the need for template matching or temporal filtering. The significance of this work lies in providing a robust, real-time detection method that reduces dependency on color information and specific shape assumptions. By eliminating the need for prior color modeling, the BCT avoids issues related to color constancy and camera calibration differences. The authors conclude that while BCT is effective for circular and polygonal signs, it is less adequate for triangular signs due to fuzzy center localization, suggesting a separate geometric model for such cases. Future work aims to integrate a recognition stage to eliminate remaining false positives from symmetrical background objects like windows or logos.

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