Document Type : Original Article
Department of Electronics, Faculty of Electrical Engineering, Shahrood University of Technology, Shahrood, Iran
Lane detection is amongst the most important operations during the automatic driving process. This process aims to detect lane lines to control the vehicle’s direction in a specific lane on the road and it can be effective in preventing accidents. Besides being online, the most requirement for lane detection methods in automatic driving is their high accuracy. The use of deep learning to create fully automated systems in lane detection has been done extensively. Automated learning-based methods used for road lane detection are often of supervised type. One of the disadvantages of these methods, despite their excellent accuracy, is that they need a set of labeled data, which limits the development procedure of the lane detection system, and most importantly establishing a standard set of labeled data is very time-demanding. The recommended solution is to use appropriate learning approaches that can be used to achieve the relative accuracy of the supervised approaches and to improve their speed. It also enables us to use different datasets without a constraint label that the tagged dataset would create in the algorithm development relative to the new dataset. In this research, we present an automatic semi-supervised learning method using deep neural networks to extract the data of lane lines, using labeled (mask for detected lines) and unlabeled datasets. The results demonstrate the suitable accuracy of the adopted method according to the proposed approach and also improves its computational complexity due to the significant reduction in the number of teachable network parameters.