TRAFFIC SIGN DETECTION USING CONVOLUTIONAL NEURAL NETOWORK ON AUTONOMOUS VEHICLE SYSTEM

• Main Author: Fijar Mayoza, Rabbi
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/50368
• Institution: Institut Teknologi Bandung

A fully automated vehicle can be applied by imitating how humans drive. Autonomous vehicle obtains this ability from artificial intelligence (AI) installed on itssystem. It is believed that AI can reduce the risk of accidents caused by human error. One example of AI implementation on autonomous vehicle is visual system for detecting traffic sign. Convolutional Neural Network (CNN), a part of deep learning method, is used in this research to build traffic sign detection model for Indonesia. However, dataset is needed by this method to perform training. The unavailability of Indonesian traffic sign dataset may become challenge in building the model due to the distinct characteristics of traffic sign among countries. The proposed solution is to feed Extended Malaysian Traffic Sign Dataset (EMTD) into CNN to produce the detection model by reason that it contains traffic signs that are similar to Indonesian traffic signs. The solution adapts Faster R-CNN model which has been developed for detecting foreign traffic sign. The CNN model is coded with Python 3 using Keras-Tensorflow library. Input data preprocessing includes resize, histogram equalization, augmentation, and pixel scaling. Experiment result shows that model with VGG16 topology gives 0.534 mAP, better than model with ResNet50 topology which gives 0.489 mAP. While ResNet50 gives better performance in process speed, 1.51 fps, than the model with VGG16 topology which is 0.34 fps.