Enhancing Perception for Autonomous Driving Systems with Dual-Vision Input Cycle GAN for Night-to-Day Image Translation

Autonomous Driving Systems (ADS) and Advanced Driver Assistance Systems (ADAS) heavily rely on vision-based data for critical decision-making. However, these systems face significant performance limitations in low-light conditions where conventional RGB vision inputs struggle to capture information due to low light. Current solutions, including high-gain cameras or LiDAR-based approaches, are expensive and increase system complexity. This study presents a novel approach to address these challenges by employing a night-to-day image translation model with modified CycleGAN. The proposed model translates RGB and infrared (IR) inputs into a daytime representation, enhancing overall perception and object detection in low-light scenarios. Unlike existing methods, the model operates in an unsupervised learning framework, avoiding the need for large, paired datasets, which are often difficult to obtain. The study also introduces architectural modifications to the CycleGAN, including a dualinput framework and refined architecture, to optimize performance across varied conditions. The model’s performance is demonstrated through reference-less evaluation metrics, achieving average scores of FID 8.04, NIQE 4.92, and BRISQUE 36.3. These results illustrate the potential of the model to enhance the safety and operational stability of autonomous driving systems in low-light environments.