Development of Real-Time Traffic-Object and Traffic-Sign Detection Models Applied for Autonomous Intelligent Vehicles

Abstract

Technologies for detecting traffic objects are an essential requirement for any applications in autonomous intelligent vehicles. In this work, models for detecting traffic objects were developed. Based on the existing datasets and the pre-trained models, fine-tuning techniques were applied to achieve trained models with higher accuracies even for the very challenging test data. The traffic object detection was developed based on the pre-trained YOLOv5s model. Two approaches were introduced for the traffic sign detection task. The so-called tiling technique incorporated with the YOLOv5s model was exploited in the first approach. In the second approach, a combination of the RetinaFace model for the localization and the MobileNetV1-SSD for the classification was employed. The experimental results show that all developed models release a very high rate of accuracy with a maximum AP50 of 75.9% for object detection and mAP50 of 64.2% for sign detection. Models developed via the second approach have twofold advantages in terms of accuracy and computational efficiency, which allows to deploy practical applications.