CNN 소실점 검출을 이용한 차선 검출

Lane detection is essential in autonomous navigation. Conventional algorithms use hand crafted features which produce difficulties because of diverse image variations from illumination variations, occlusions and shadows. Recently, deep learning based approaches have provided more robust results. In this paper, we present an algorithm for the robust detection of lanes by finding vanishing points with convolutional neural networks. We use two modified CNN architectures, where the final output layer consists of four elements. The epipole and the angles of the current driving lane each have two elements. Experiments are performed by using two modified structures of the NVIDIA end-to-end model[9] and the ResNet-50 model[10].