METHOD TO IMPROVE THE YOLO-BASED TRAFFIC SIGN IDENTIFICATION SYSTEM PHƯƠNG PHÁP CẢI TIẾN HỆ THỐNG NHẬN DẠNG BIỂN GIAO THÔNG BẰNG ỨNG DỤNG YOLO

Abstract

With the rapid development of intelligent transportation, more and more vehicles are equipped with intelligent traffic sign recognition systems, which can reduce potential safety hazards caused by human perception errors. Therefore, a safe and reliable traffic sign recognition system is not only a requirement for drivers but also a hot research topic for automobile manufacturers. To meet this demand, the authors propose in this paper a traffic sign recognition system based on computer vision techniques and You Only Look Once (YOLO) algorithm. The system is designed to recognize four different types of traffic signs including: no left turn, no right turn, no stopping or parking, and no parking. In order to train the proposed system, images were collected from roads in Ho Chi Minh City including 735 images of signs prohibiting left turns, 713 images of signs prohibiting right turns, 177 images prohibiting left-right turns, 752 images no stopping-no parking signs, 629 photos of no parking signs, 191 photos of prohibiting cars from turning right, 143 photos of prohibiting cars from turning left, 171 prohibiting turning around and 109 prohibiting going straight. The system was then experimentally tested in the field for recognition accuracy according to the mAP@.5 measure as follows: 99.4% correctly identified signs prohibiting left turns, 99.3% correctly identified signs prohibiting left turns right, 95.6% correctly identified no-stop-no-parking signs, 95.3% correctly identified no-parking signs, 98.6% correctly identified no-turn signs, 93.7% correctly identified no-turn right signs, 94.5% correctly identified Correct signs prohibiting cars from turning left, 99.5% correctly recognized signs prohibiting left-right turns and finally 93.4% correctly identified signs prohibiting going straight.