Content area
Full Text
Abstract
Traffic-sign recognition is critical for vehicle safety applications, especially as self-driving cars become a reality. This paper proposes a solution based on existing approaches, utilizing deep learning and computer vision preprocessing to create a real-time algorithm that addresses the limitations of previous methods. The proposed algorithm aims to overcome as many drawbacks as possible and serve as a core component of advanced driver-assistance systems (ADAS). The proposed method is evaluated using the German Traffic Sign Recognition Benchmark (GTSRB) and the Belgium Traffic Sign Dataset (BTSD). This study concludes with a fully functional pipeline that can inspire the development of driving assistants and advance the future of self-driving cars.
Keywords: traffic-sign recognition, detection, neural networks, deep learning
JEL Classification: C45, C88
1. Introduction
According to the World Health Organization's statistic on road traffic injuries [1], 1.3 million people die annually due to road crashes, with 93% of these crashes occurring in less developed countries where older cars with less equipped safety technology are in circulation. Additionally, between 20 to 50 million people get injured each year, leading to a 3% cost of a country's gross domestic product. This highlights the importance of detection and recognition of road signs in the real world. The European New Car Assessment Programme (Euro NCAP) has made it mandatory for all new cars sold in the EU to be equipped with this type of technology soon [2], as they place great value on car safety and have conducted surveys and safety campaigns regarding ADAS, stating that cars of the future need "readable" roads [3]. Detecting different signs in varying weather, daytime and road conditions is seen as a challenge, as current advanced driver assistance systems only have a defined subset of possible signs. Unfortunately, no comprehensive unbiased comparison of sign detecting algorithms has been implemented, and the slow development of this feature may be attributable to a lack of a big freely available benchmark data set.
The recognition process can be divided into two steps: detection and classification. When comparing the two, detection takes precedence because state-of-the-art classification systems can only compete with humans at best. Therefore, classification can be regarded as solved, at least for the time being [4][5]. While most of the attention of sign detection is...