Real-Time Application of Traffic Sign Recognition Algorithm with Deep Learning

Abstract

Autonomous vehicles are one of the increasingly widespread application areas in automotive technology. These vehicles show significant potential in improving transportation systems, with their ability to communicate, coordinate and drive autonomously. These vehicles, which move from source to destination without human intervention, appear to be a solution to various problems caused by people in traffic, such as accidents and traffic jams. Traffic accidents and traffic jams are largely due to driver faults and non-compliance with traffic rules. For this reason, it is predicted that integrating artificial intelligence (AI)-based systems into autonomous vehicles will be a solution to such situations, which are seen as a problem in social life. Looking at the literature, VGGNet, ResNet50, MobileNetV2, NASNetMobile, Feed Forward Neural Networks, Recurrent Neural Networks, Long-Short Term Memory, and Gate Recurrent Units It is seen that deep learning models such as these are widely used in traffic sign classification studies. Unlike previous studies, in this study, a deep learning application was made for the detection of traffic signs and markers using an open-source data set and models of YOLOv5 versions. The original data set was prepared and used in the study. Labeling of this data set in accordance with different AI models has been completed. In the developed CNN models, the training process of the data set containing 15 different traffic sign classes was carried out. The results of these models were systematically compared, and optimum performance values were obtained from the models with hyperparameter changes. Real-time application was made using the YOLOv5s model. As a result, a success rate of 98-99% was achieved.