Enhancing Object Recognition for Self-Driving Cars: Rainfall Data Generation Using CycleGAN
Yejin Ha, Yi Chang, Kyo-Chan Koo, Jae Hyung Cho · Asia-pacific Journal of Convergent Research Interchange · 2025
Autonomous driving systems require extensive visual data, particularly for challenging conditions like heavy rain.Access to rainy weather image data is essential for training object recognition systems to perform effectively in real-world rainy conditions.However, collecting rainy weather data is more time-consuming and expensive than gathering data in normal conditions, resulting in an unbalanced dataset.To address this issue, this study proposes using data augmentation techniques to generate rainy weather data.CycleGAN was employed to generate synthetic rainy weather images.This approach enabled the creation of a substantial amount of training data.The model was trained using 5,000 randomly selected normal weather images and 5,000 rainy weather images for ten epochs.This process was repeated ten times, resulting in 100 epochs.The trained CycleGAN model generated 105,305 synthetic rainy weather images from an equal number of normal weather images.The primary goal of this data augmentation is to improve the autonomous driving system's object recognition.Object recognition experiments using the YOLOv5 model were conducted to evaluate its effectiveness.Three sets of learning data are used in these tests for object recognition: a small set of data with rain, a big set of data with normal weather, and supplemented rainy weather data with CycleGAN.The average accuracy, as indicated by the improvement in mAP to 0.355, demonstrates the practical benefit of augmenting training data using CycleGAN, enabling enhanced recognition accuracy in rainy conditions, which is essential for real-world autonomous vehicle performance.This result highlights the effectiveness of CycleGAN-based augmentation in improving object recognition with specialized data.This methodology successfully tackles the problem of data imbalance resulting from data augmentation and demonstrates its effectiveness in producing training of superior quality.