Smart data driven traffic sign detection method based on adaptive color threshold and shape symmetry

Highlights

• Robust and adaptive traffic sign detecting algorithms in complex traffic environment.

• Adaptive threshold traffic sign segmentation by approximate max-min normalization.

• High efficient geometrical symmetry detection with statistical hypothesis testing.

• The detection rate reached above 94% on the German traffic sign dataset GTSDB.

Abstract

Detection and recognition of road traffic signs constitute an important element in Advanced Driver Assistance Systems (ADAS), which can provide real-time road sign perception information to vehicles. In this paper, we proposed a new traffic sign detection method based on adaptive color threshold segmentation and the hypothesis testing of shape symmetry by leveraging traffic signs and image data. First, we calculated an adaptive segmentation threshold using the cumulative distribution function of the image histogram. Based on this, we designed an approximate maximum and minimum normalization method, which is used to suppress the interference of high brightness area and background in image thresholding processes. Secondly, we transformed the highlight shape feature of thresholding image into a connected domain feature vector. And we formulated a shape symmetry detection algorithm based on statistical hypothesis testing to efficiently extract the ROI of traffic signs based on traffic data analysis. We performed some comprehensive experiments on the GTSDB (German Traffic Sign Detection Benchmark) dataset. The accuracy of traffic sign detection exceeded 94%. This method has higher detection accuracy and time efficiency than other methods, and better robustness under complex traffic environment.

Introduction

Detection and recognition of road traffic sign constitute an important technological element in Advanced Driver Assistance Systems [1] (ADAS), which can provide real-time road sign information to vehicles and enhance driving safety [2]. Traffic signs are usually composed of specific shapes (round, square, and triangular) and colors (red, blue, and yellow), which have significant visual effect in the road environment. Therefore, traffic sign detection can be classified into color-based, shape-based and color-shape-based methods [3], [4].

In color-based traffic sign detection, the RGB images are usually transformed into other color space such as HSI [5], CIELab [6], and HSL [7]. Then the traffic signs are extracted by color thresholding segmentation based on smart data processing. Color-based traffic sign detection methods are easily affected by complex illumination conditions of traffic scenes. In shape-based traffic sign detection, geometric contour shapes of traffic signs are detected by means of template matching [8], geometric invariant moment [9], and geometric symmetry [6], [10]. Symmetry detection has better adaptability compared with template matching, geometric invariant moment in a complex illumination environment, but it requires higher computational complexity.

The existing color and shape based traffic sign detection methods have weaker adaptability under complex brightness conditions. In this paper, we propose an adaptive color threshold segmentation and high efficient shape symmetry algorithm to achieve robust traffic sign detection in a complex illumination environment. (1) In color thresholding segmentation: we calculate an adaptive threshold using the cumulative distribution function of image histogram; and we use the approximate maximum–minimum normalization method to process images according to the threshold. With this method, the interference of overexposure foreground objects and background is suppressed, so we can achieve better traffic sign segment effect in complex lighting environments. (2) In geometrical symmetry detection: by converting connected regions of a binary image into shape geometrical feature vectors, we use the statistical hypothesis testing method to judge the symmetry of connected components in a binary image, and extract symmetry components as the candidate ROIs of traffic signs. With this method, we realize higher efficient geometrical symmetry computation of the image than with the previous algorithm based on frequency domain.

The traffic sign detection process is shown in Fig. 1. The main contributions of this paper are included in the second and fifth steps (detailed algorithms are presented in Section 3).

Step 1: Traffic images are transformed into Red–Blue gray images by Red–Blue normalization processing [11];

Step 2: The adaptive segmentation threshold is calculated based on the cumulative distribution function of the histogram of Red–Blue images; and the images’ foreground enhancement is processed using the approximate maximum–minimum normalization method to suppress the images’ high brightness background;

Step 3: Candidate connected domains of traffic signs are extracted by morphological filtering and MSER [11];

Step 4: Enhanced processing of the connected domains is performed using convex operation;

Step 5: Symmetry statistical analysis and hypothesis testing are performed on the feature vectors of connected domains to detect shape symmetry and extract candidate ROIs of traffic signs.

Step 6: Traffic signs’ ROI is determined using geometric constraints.

The paper is organized as follows: Section 2 describes related works about traffic sign detection, Section 3 details the main contributions of this paper, Section 4 shows some experimental results of our method, and Section 5 presents the conclusions.