International Journal of Computer Applications (0975 – 8887) Volume 84 – No 15, December 2013 32 Speed Sign Recognition using Shape-based Features Jafar Abukhait Imad Zyout Ayman M Mansour Tafila Technical University Tafila Technical University Tafila Technical University P. O. Box 179 P. O. Box 179 P. O. Box 179 Tafila 66110, Jordan Tafila 66110, Jordan Tafila 66110, Jordan ABSTRACT An efficient shape-based recognition system of U.S. speed limit road signs is presented in this paper. The proposed system accomplishes speed sign detection and recognition processes using three main stages, namely, geometrical-based detection of rectangular road signs, shape-based segmentation and feature extraction, and pattern classification using a K-nearest neighbor classifier (KNN). Twenty shape descriptors are computed for the most discriminative numerals of each detected sign. The proposed system is invariant to scale, rotation, and partial occlusion. The proposed system has been tested in different conditions, including sunny, cloudy, and rainy weather, and the experimental results on 195 speed signs reveals the efficiency of the proposed shape pattern segmentation and feature extraction methods. General Terms Speed Sign Recognition, Morphology-based Features, Feature Extraction, Classification. 1. INTRODUCTION Driver assistant systems (DAS) and autonomous vehicles have been suggested by transportation departments due to high investment in road network management. The goal of such systems is to make driving safer and more reliable [1]. Such systems can benefit from road signs information by applying artificial intelligence and computer vision techniques. Nowadays, automated road sign detection and recognition systems are being proposed since road signs have a significant role in regulating traffic and warning drivers over highways. Detection and recognition are considered as two different and cascaded problems. The purpose of detection is to maintain sign objects and remove non-sign objects from road sign frames while recognition is used to classify different road signs to their categories. In general, detection of road signs is accomplished using: color- based, shape-based, or integrated methods that use both color and shape information simultaneously. Sign’s color information is used in color-based methods to remove non-road sign objects from the scene. While color thresholding in RGB is being used to segment road sign images ([2], [3], [4]), other researchers use Hue Saturation Intensity (HSI) space in the segmentation process ([5], [6]). Shape-based detection methods are being deployed to detect road signs objects including speed sign ones. In [7, 8], distance to border (DtB) vector is used to build the shape feature vector which is used in shape classification by SVM, while principal component analysis (PCA) and k-nearest neighbor (KNN) classifier are used to detect the sign in [4]. Hough transform and radial symmetry are used to recognize triangular and circular shape road signs in [9]. Haar-like features are used in [10] while Genetic algorithm is used in [5] to detect road sign shape. In [13], fast radial symmetry is used to detect circular speed signs from a gray-scale image without using color information. In [14], a self-organizing map (SOM) is used on Gabor wavelet convoluted images to distinguish road signs from non-road signs. In the classification, different techniques were suggested to improve classification accuracy and processing time. Many road sign recognition systems use Artificial Neural Network (ANN) in the classification process ([5], [6], [15], [16], [17]). Template matching using cross correlation is also used to identify road sign objects ([4], [8], [18], [19]). Statistical, structural, or spectral features along with different classifiers are used in road sign recognition. Histogram is used as an image descriptor in [20] while Principal Component Analysis (PCA) is used in [4]. Zernike moment [21], Scale- invariant feature transform (SIFT) [22], and color distance transform (CDT) ([23], [24]) are also deployed as image descriptors. Support vector machine (SVM) ([2], [21]), K- nearest neighbor (KNN) [4, 20, 22], Forest Error-Correcting Output Code (F-ECOC) [9], and Bayesian generative modeling [25] are used as classifiers in the road sign recognition process. This work is intended to implement an automated system for the recognition of speed limit signs in the United States. The proposed system has the ability to classify six different categories of rectangular shape speed signs: Speed Limit 15, Speed Limit 25, Speed Limit 35, Speed Limit 45, Speed Limit 55, and others (which include all rectangular white signs other than the previous five speed signs) as shown in Figure 1. The proposed system, automatically, segments the numerals of the speed sign using morphological operations; then different shape-based features of these numerals are extracted to build a feature vector that is used by a KNN classifier. SPEED LIMIT 15 SPEED LIMIT 55 SPEED LIMIT 45 SPEED LIMIT 35 SPEED LIMIT 25 PASS WITH CARE Figure 1: The six sign categories tested by our proposed system. 2. PROPOSED SHAPE-BASED SIGN RECOGNITION SYSTEM The proposed automated detection and recognition system of U.S. speed signs, as shown in Figure 2, follows five stages. 1. Color-based segmentation: speed sign objects are extracted by converting sign images to achromatic ones and applying RGB color thresholding. 2. Rectangular speed sign shape detection: this stage discard non-speed sign objects and detects rectangular shapes using a set of cascaded geometric detectors.