Traffic Lights and Traffic Signs Detection System Using Modified You Only Look Once

Abstract

Traditional image processing methods used for detecting traffic lights and traffic signs are replaced by the recent enhancements of the deep learning method by the success of building a Convolutional Neural Network (CNN). In this research, a traffic lights and traffic signs detection system using a modified You Only Look Once (YOLO) has been proposed. The system processes an image captured by a camera sensor and provides the results in the form of detecting traffic lights and traffic signs contained in the image. The CNN architecture used is a modified Cross Stage Partial YOLOv4 (YOLOv4-CSP). The experiments were carried out using a self-constructed dataset consisting of1360 training data and 340 testing data with 6 types of traffic lights and 39 types of traffic signs. The network is built using the Darknet framework and the result shows 79,77% of the mean Average Precision at the 0,5 Intersection over Union threshold (mAP at 0,5 IoU threshold) and 29 frames per second (FPS) of inference speed tested on a single NVIDIA Tesla T4 Graphics Processing Unit (GPU).