Normalizing images in various weather and lighting conditions using Pix2Pix GAN

Abstract

Autonomous vehicles are widely regarded as the future of transportation due to its possible uses in a myriad of applications. In recent years, perception systems in driverless cars have had reasonable development through the various implementations of object detection systems with deep-learning algorithms. Noticeable progress has been made in this field of study as many isolated and multi-model systems have been developed and/or proposed to help overcome the shortcomings of the sensors and detection algorithms. These include research on sensing objects under varying environmental conditions (illumination, refractive indexes, weather conditions) as well as detection and removal of noise, clutter, and camouflage from the collected sensory inputs. However, in its current state, perception systems in autonomous vehicles are still incapable of accurately detecting objects in real-life scenarios using its visual/thermal camera, LiDAR, radar, and other sensors. Additionally, most systems lack the robustness to perform well under any given condition. Hence, this paper proposes to use advanced color vision techniques and Generative Adversarial Networks (GAN) to produce reconstructed images that can improve the accuracy of object detection systems for more precise predictions.