준지도 학습을 위한 신호등 학습 데이터 선별

초록

To enable autonomous driving in urban areas without V2X infrastructure, traffic light recognition is essential. This paper addresses the problem of decreased recognition performance due to changes in illumination on the road. To resolve this issue and enhance learning efficiency with limited datasets, we propose an image dataset selection algorithm based on brightness and sharpness to include diverse environments. Additionally, we integrate datasets to recognize both horizontal and vertical traffic lights with a single model by redefining the classes. Experiments demonstrate stable detection and recognition on both American and Korean datasets. In a semi-supervised learning environment, starting with partially labeled datasets, the proposed dataset selection algorithm outperforms random and feature-based selection methods in recognition performance after training.