TLGAN: Conditional Style-Based Traffic light Generation with Generative Adversarial Networks

Traffic light recognition plays a vital role in intelligent transportation systems and is a critical perception module for autonomous vehicles. Compared with cars, pedestrians, and other targets, the traffic light has the characteristics of variety and complexity and their state constantly changing, which adds many difficulties to recognition. The performance of the entire deep learning-based vision system largely depends on whether its training dataset is rich in scenes. However, it is difficult to collect data in rare scenarios such as extreme weather, flashing, and no working, resulting in data imbalance and poor model generalization ability. This paper proposes a model called TL-GAN, a conditional style-based generative adversarial network, to generate images of traffic lights that we lack, especially yellow, inactive and flashing traffic lights. Our model uses style mixing to separate the background and the foreground of the traffic light and apply a new template loss to force the model to generate traffic light images with the same background but belonging to different classes. In order to verify the validity of the generated data, we use a traffic light classification model based on time series. The results of experiments show that AP(average precision) values of the three categories have been improved by adding generated images, proving the generated data's validity.