End-To-End Ego Lane Estimation Based on Sequential Transfer Learning for Self-Driving Cars

Autonomous cars establish driving strategies using the positions of ego lanes. The previous methods detect lane points and select ego lanes with heuristic and complex postprocessing with strong geometric assumptions. We propose a sequential end-to-end transfer learning method to estimate left and right ego lanes directly and separately without any postprocessing. We redefined a point-detection problem as a region-segmentation problem; as a result, the proposed method is insensitive to occlusions and variations of environmental conditions, because it considers the entire content of an input image during training. Also, we constructed an extensive dataset that is suitable for a deep neural network training by collecting a variety of road conditions, annotating ego lanes, and augmenting them systematically. The proposed method demonstrated improved accuracy and stability on input variations compared with a recent method based on deep learning. Our approach does not involve postprocessing, and is therefore flexible to change of target domain.