Makine ile öğrenme yöntemleriyle trafik işareti tanıma

Traffic sign recognition is an important subject for the vehicles that are autonomously controlled. Since there is no human intervention in this type of vehicle, all information about the environment is collected by means of sensors and systems such as camera, distance sensor, RFID, sonar sensor and GPS. One of the important requirements for autonomous vehicles is the recognition of traffic signs. In the first methods proposed for feature extraction in the identification of traffic signs was generally performed by computer vision method, and classification was performed by using these attributes, but due to the large and fast increase in GPU performance, feature extraction was done by machine learning methods. CNN has become one of the most used deep learning method thanks to its state-of-the art performance at tough problems. With CNN, superior success has been achieved in traffic sign recognition which is crucial for autonomous vehicles. In this study, two-stage hierarchical CNN structure is proposed in order to classify traffic signs. Signs are divided into 4 main groups at the first stage by using different similarity indexes. And then classes of each main group are subclassified with CNNs at the second stage. Performance of the network is examined on 43-classes GTSRB dataset, compared with structures using different similarity index and methods proposed in thesis.