심층 컨볼루션 신경망을 이용한 도로 차선 모델 기반 자율 차로유지 제어시스템

Abstract

카메라 기반 자율 차로유지 시스템(Lane Keeping System, LKS)은 차로의 중심선을 따라 주행하도록 설계되어 있어 신뢰할 수 있는 도로 차선의 정보의 획득이 매우 중요하다. 그러나, 기존의 머신 비전(Machine Vision) 시스템은 신뢰할 수 있는 차선 정보, 즉 도로 계수를 제공하지 못하는 경우가 빈번하다. 본 논문에서는 이러한 문제를 해결하기 위하여 심층 컨볼루션 신경망(Deep Convolutional Neural Network, DCNN)을 이용하여 전방 카메라로부터 가상 차로의 도로 계수를 추출하는 방법을 제안한다. DCNN은 운전자가 수동으로 주행한 실험 차량 내부에 탑재된 비전 시스템이 출력한 도로 계수와 다른 카메라로부터 얻어진 이미지로 구성된 데이터 세트를 활용하여 학습되었다. 제안된 DCNN의 실효성을 확인하기 위해, 학습할 때 사용되지 않은 세 가지 시나리오의 데이터 세트를 활용하여 제안된 방법의 유효성을 다중 주기 상태추정를 이용한 LKS를 구축하고 검증하였으며 각 시나리오에서의 정량적 오차 통계 분석을 수행하였다. 이를 통해, 운전자가 주행 중 입력한 조향각과 제안한 LKS의 제어 입력인 조향각 사이에서 높은 상관성을 확인하였다. 차선이 없거나 차량이 교차로에서 선회하는 특정 상황 등에서

도 제안된 DCNN은 가상 도로 차선 모델의 계수를 출력하는 것을 관찰했다.|It is very important to obtain reliable coefficients of a road lane information because the vision-based autonomous lane keeping system (LKS) is designed to track the centerline of the road lane. However, a conventional machine vision system often fails to provide the reliable lane information, that is, road coefficients. In this paper, we propose a novel method for extracting the road coefficients for a virtual lane using a deep convolutional neural network (DCNN) to solve these problems. The DCNN was trained with a dataset consisting of reliable road coefficients from the vision system mounted on the test vehicle driven by a human, and captured raw images by another heterogenous camera. For validation, we utilized a dataset not used for training, which has three scenarios. Performance of the proposed method was evaluated with statistical analysis of error with the validation dataset. We confirmed that there were good agreements between steering wheel angles by the human driver and those given by the proposed LKS. Furthermore, we observed that the proposed system can provide the road coefficients even either when there is no lane marker and/or when the vehicle is maneuvered for turning at an intersection.; It is very important to obtain reliable coefficients of a road lane information because the vision-based autonomous lane keeping system (LKS) is designed to track the centerline of the road lane. However, a conventional machine vision system often fails to provide the reliable lane information, that is, road coefficients. In this paper, we propose a novel method for extracting the road coefficients for a virtual lane using a deep convolutional neural network (DCNN) to solve these problems. The DCNN was trained with a dataset consisting of reliable road coefficients from the vision system mounted on the test vehicle driven by a human, and captured raw images by another heterogenous camera. For validation, we utilized a dataset not used for training, which has three scenarios. Performance of the proposed method was evaluated with statistical analysis of error with the validation dataset. We confirmed that there were good agreements between steering wheel angles by the human driver and those given by the proposed LKS. Furthermore, we observed that the proposed system can provide the road coefficients even either when there is no lane marker and/or when the vehicle is maneuvered for turning at an intersection.