Detection and recognition of signs is an automobile equipment that reads and interprets permanent and temporary signs located at the edge or over the road, in order to inform the driver in case he could not see them. Speed limitation and over-ride signs are particularly concerned. Signs detection and recognition works through a camera mounted behind the interior rearview mirror. It detects signs located the left, right or over the road and compares it with an internal dataset. Once the sign is recognized, the driver is notified of the situation through a visual on the GPS or instrumentation. It is a useful aid for the driver who, in the ambient traffic, will not necessarily have seen the sign. Depending on the system, the car goes up to compare its speed with the current limitation and alert the driver if he is overspending. In the long term, we can also imagine that the optical reading of the signs can communicate with the adaptive cruise control. It is the vehicle that would automatically decide how fast to drive.
Traffic Sign Detection and Recognition (TSDR) has been very popular in recent years. This is due to the large number of applications that such a system can provide:
A. Road Sign Detection:
Methods of detection of road signs are divided into three classes. Methods based on color, shape or machine learning. The dominant colors of most road signs are red, blue or yellow. Many authors (Lillo-Castellano, et al., 2015; Ellahyani, et al., 2016) use this property to detect signs. An associated component segmentation based on a color model is used. The regions of interest are then validated by a recognition algorithm or an appearance model. These methods are usually fast and invariant to translation, rotation, and scaling. Since color can be easily affected by lighting conditions, the main difficulty of color-based methods is how to be invariant to different lighting conditions. These methods tend to follow a common pattern: the image is transformed into a color space and then thresholded. The two spaces HSV and HSI are very used by researchers because they are based on human perception of colors and encode color information in one channel instead of three (Ardianto, et al., 2017).
For shape-based methods, the contours of the image are analyzed by a structural or global approach. These methods are generally more robust than color-based ones because they handle the gradient of the image and can handle grayscale images. These methods are very sensible to occlusions and deformation, that affects considerably their performances. To overcome this problem some researchers proposed to detect circular road signs using the circle detection algorithm EDCircles (Kaplan, et al., 2016) and other authors proposed to detect circular and triangular signs using HOG and Linear SVM (Zaklouta and Stanciulescu, 2014). systems that adopt shape-based methods to minimize color change due to lighting and climate change face the problem of detecting occluded and damaged signs that require color-based methods.
Finally, for methods based on machine learning, a classifier (cascade, SVM, neural networks) is trained based on examples. It is applied on a sliding window that traverses the image on several scales. These methods combine geometry and photometry but can be a costly step in computing time. They require the constitution of a learning base by type of signs, tedious step when the number of objects to be detected is large. Many researchers (Ellahyani, et al., 2016; Brkić, et al., 2010; Chen and Lu, 2016; Yi, et al., 2016; Bouti, et al., 2017) adopt this approach of combination between color-based methods and shape-based methods that can help to minimize the rate of false positives and increase the rate of true positives.