Abstract:

Today, artificial intelligence has become an indispensable part of our lives with the constantly advancing technological developments. One of the fields of study where artificial neural networks are used is transportation. Artificial intelligence is used in driver support systems applications in order to reduce possible accidents in the field of transportation. In this study, a dataset consisting of images of 4000 traffic sign images in total, both the dataset created individually by taking the photos of the traffic signs and the dataset obtained from the open source website (kaggle.com) was used. The dataset contains 3200 training data and 800 test data. The prepared data sets were trained with three different deep learning methods, namely ResNet50, MobileNetV2 and NASNetMobile, together with the CNN (Convolutional Neural Network) model, and their performance was evaluated according to the factors of training accuracy, test accuracy, training loss and test loss. With the ResNet50 method, the training accuracy was 97.62%, the test accuracy was 78.75%, the training loss was 0.1%, and the test loss was 6.28%. With the MobileNetV2 method, the training accuracy was 97.8%, the test accuracy was 48.12%, the training loss was 0.38%, and the test loss was 38.34%. With the NASNetMobile method, the training accuracy was 98.56%, the test accuracy was 41.56%, the training loss was 0.1%, and the test loss was 17.28%.

Keywords: