Improving the Performance of Universal Traffic Light Recognition Through Dataset Construction and Selection in Semi-supervised Learning

Traffic light recognition in autonomous driving is an essential but very challenging task because its performance is affected by unpredictable environmental conditions. Moreover, the shapes and installations of traffic lights in various countries require a lot of cost and time to collect a huge dataset and train a deep learning model. This study develops a method for constructing training datasets with minimal resource expenditure for more efficient and universally applicable traffic light recognition. It aims to develop a deep learning scheme that can universally recognize various international traffic light configurations by replacing the initial supervised learning stage with a phase that can recognize multiple types of traffic lights and improve performance through self-supervised learning on new datasets. Additionally, this study introduces a data selection algorithm that enables robust recognition under changes in road conditions by enhancing adaptability across diverse environments. This research not only demonstrates improved recognition performance in hazardous road situations but also concludes with potential for efficient application worldwide. Experimental results showed 24.2% improvement in mAP50, even with the same proportion of labeled data, and similar performance with half the labeled data.