Traffic Sign Detection Using YOLOv3

Abstract

Advanced driving assistance systems (ADASs) are increasingly being installed in modern vehicles because they make driving safer and more comfortable. With the implementation of cameras in the vehicle, the range of possible ADASs increases. One of such systems is the one aimed for traffic sign recognition, which alerts the driver about different road conditions such as excess of the speed limit or traffic ban. In this paper, a solution for detecting a specific set of 11 traffic signs typical for most European countries is presented. The algorithm used for detecting traffic signs is You Only Look Once (YOLO) v3, where the model parameters are trained on a train set acquired from the newly created dataset. The rest of the dataset images are used for creating a test set. The dataset is derived from the video signals that were capturing traffic with a front view camera mounted inside the vehicle, in the city of Osijek in different weather conditions (sunny, cloudy, rain, night). The dataset images are extracted from 28 different video sequences, which resulted in 5567 images with the total number of 6751 annotated traffic signs. The proposed solution for detecting a specific set of traffic signs achieves high performance when tested on the test set created from the proposed dataset.