Article Preview
TopIntroduction
Traffic sign recognition (TSR) has a vital importance in autonomous driver assistant system (Stallkamp, Schlipsing, Salmen, & Igel, 2012). The detection and recognition of traffic signs and symbols from digital images have been an area for research over last few years. Two major tasks are generally involved in a typical TSR system: detecting the approximate sizes and locations of traffic signs on the road (traffic sign detection-TSD) and then categorizing those detected traffic sign’s images into their respective classes (TSR). Traffic sign detection and recognition system (TSDR) is a driver supportive system may also be used to inform and warn the driver in difficult circumstances. This system has the ability to detect and recognize all traffic signs, even those signs that may be occluded or somewhat distorted.
TSD is the initial phase of any TSDR system. In this phase, candidate region of interests (ROIs), which is the area with most probability to have the regions of traffic signs, are detected. TSD involves locating potential sign image regions from a natural scene image input. This initial stage is called the detection stage, in which a ROI-containing traffic sign is actually localized (Møgelmose, Trivedi & Moeslund, 2012; Gündüz, Kaplan & Günal, 2013). Traffic signs usually are of red, blue or white color and specific shapes like round, square, and triangular. These characteristics differentiate the traffic signs from other outdoor objects, which makes them appropriate to be computed by a computer vision system. This allows the TSDR system to distinguish traffic signs from the background scene (Yuan, Xiong, & Wang, 2017). Thus, the traditional detection method consists of either based on color or shape or a hybrid of both.
Traffic signs are displayed with standard shapes and attractive colors that appeals the human drivers’ attention. Various machine learning (ML) and deep learning (DL) methods viz. support vector machines (SVM) (Agrawal & Chaurasiya, 2017; Do et al., 2017; Kiran, Prabhu, & Rajeev, 2009; Soni, Chaurasiya, & Agrawal, 2019), k-nearest neighbor (KNN) classifier (Sugiharto & Harjoko, 2016), fully convolutional network (FCN) (Y. Zhu et al., 2016), and convolutional neural network (CNN) (Aghdam, Heravi, & Puig, 2016; Farag & Saleh, 2018; Z. Zhu et al., 2016) have been employed for classification in TSR system.
There has been a lot of research in the field of TSR for the past few years. Jin et al. have designed a CNN-based model achieving high accuracy using a hinge loss stochastic gradient descent method (Jin, Kun Fu, & Changshui, 2014). A multi-column deep neural network for traffic sign classification was proposed by CireşAn et al., where they proclaimed better results than human-recognition-rate (CireşAn et al., 2012),. A fast template matching technique was proposed by Torresen et al. (Torresen, Bakke, & Sekanina, 2004). Li et al. applied adaboost learning with five classical Haar wavelets and four histogram of oriented gradient (HoG) features (Li, Pankanti, & Guan, 2010). Youssef et al. proposed a combination of color segmentation, HoG, and CNN in their TSDR system to achieve improved classification accuracy and computational speed (Youssef et al., 2016). Qian et al. used CNN to learn the discriminative feature of max pooling positions to classify traffic signs and obtained a comparable performance with the-state-of-the-art methods (Qian et al., 2016). Greenhalgh and Mirmehdi depicted a comparison between different classifying techniques namely SVM, MLP, HoG-based classifiers, and decision trees. They found that decision tree with average accuracy of 94.2% had the highest accuracy with lowest computational time (Greenhalgh and Mirmehdi, 2012).
German traffic sign recognition benchmark (GTSRB) dataset has been widely used by researches for testing their proposed approaches (Stallkamp, Schlipsing, Salmen, & Igel, 2011). Most of traffic sign images detected from GTSRB dataset are of very low quality. Hence, pre-processing and normalization of the images has helped to improve the quality of them. Methods used for image processing include gamma correction (Rahman, Rahman, Abdullah-Al-Wadud, Al-Quaderi, & Shoyaib, 2016), histogram equalization (Sepasian, Balachandran, & Mares, 2008), and contrast limited adaptive histogram equalization (CLAHE) (Sepasian et al., 2008).