적대적 생성 네트워크를 기반으로 추출된 불변 특징을 이용한 차량 주행 방법 학습

Abstract

As deep learning evolves, camera information becomes increasingly important in autonomous driving system. However, camera measurements are sensitive to weather or illuminance and it leads to incorrect information of camera. To design a safe autonomous vehicle system, perception algorithm which is robust to environmental changes should be designed. This paper proposes a method that extracts environment-invariant features using generative adversarial networks (GAN). The braking scenario is conducted to show the effectiveness of the extracted features through GAN. The braking policy is trained by using reinforcement learning algorithm. The network and simulation environments are implemented by using Tensorflow and Unity respectively.