Analysis and implementation of an efficient traffic sign recognition based on YOLO and SIFT for Turtlebot3 Robot

Abstract

Traffic Sign Recognition (TSR) is one of the key aspects for autonomous driving and it plays a vital role to make autonomous driving successful, but that’s only possible if TSR is efficient enough and reliable. This work addresses exploration of simple and fast to implement options for robotic applications. For analysis and implementation, we are focusing on a Turtlebot3 Robot (TB3). Various potential TSR algorithms are evaluated in different test-cases with the goal of developing an optimized TSR with accurate results for German traffic signs. Therefore, the robot was tested on its own Mini-City track. On this Track we started to detect the signs with a simple Scale-Invariant Feature Transform (SIFT). However, the accuracy of SIFT was showing limitations for the use within TSR on mini-city-Track. This approach focuses on educational use where limitations and simple applications of autonomous driving are investigated. A review of state-of-art algorithms was done, to evaluate and improve accuracy. For example, Oriented FAST and Rotated Brief algorithm (ORB), You Only Look Once (YOLO) and SIFT algorithm was tested on TB3 in a way that all important criteria are fulfilled along with system being real-time. Regarding YOLOv8 a custom dataset and training is performed. The YOLO-model achieves 99.5% in terms of mean Average Perception (mAP@0.5) for all classes. In summary, as a powerful alternative to work with, YOLOv8 was identified. Standalone or in combination with SIFT a TSR system is shown which can work impacted by several environmental conditions. Based on evaluation of three algorithms an optimized code was developed in which YOLOv8 and SIFT were used in combination as a well performing TSR algorithm, which has above 95% accuracy for each traffic sign tested.