Convolutional Neural Networks for Traffic Signs Recognition

The application fields of traffic signs recognition are multiple, including autonomous vehicles, self-driving cars, Advanced Driver Assistance Systems (ADAS), etc. The ultimate goal of these systems is to save lives on roads by reducing the number of accidents. Specifically, the incidents caused by drivers’ distraction and inattention. Efficient traffic signs recognition systems represent then one of the major components in making cars safer, not only for drivers, but for pedestrians as well. In this domain of research, Deep Learning approaches have gained more popularity due to their appealing performances. However, these approaches still face many difficulties, especially, those related to adverse conditions and real time response. Furthermore, these methods are very demanding in terms of hardware requirements, computational load and training data. From this perspective, our work aims to create an efficient Deep Learning model, for the classification of traffic signs, under adverse and challenging conditions. The objective of the research is to reach a high recognition accuracy, via a faster learning process, while using limited data and hardware resources. The obtained results show that CNNs could reach an accuracy of more than 99.6% for signs with no challenges. However, their accuracy decreases when dealing with more challenging traffic signs. The results shows also that lighter architectures are efficient in terms of generalization, accuracy and speed. © 2020, Springer Nature Switzerland AG.