Research On Lane-line Detection Method Based On Convolutional Neural Networks

Lane perception(detection)is the crucial problem in Advanced Driver Assistance Systems(ADAS),as it either alerts the driver in dangerous situations(Lane Departure Warning)or takes an active part in the driving.Thus the detection method has been an active field with considerable progress of research in the last few years.With the rapid evolution of deep neural networks,like other computer-vision tasks,methodologies of lane detection have been taken over by deep-learning in recent decades and state-of-the-art results have been superior to traditional methods,especially in complicated urban-road backgrounds.Current issues in lane detection are:Firstly,characteristics of lanes may be inconspicuous as fading effect;Then shade and building may interferes detection result;What’s More,detailed information may be lost as lane pixels are markedly small compared to the whole picture;Last but not least,occlusions commonly interfere in the detection.This paper contributes several methods corresponding to the above problems:(1)A context extraction method is proposed.This method gains more discriminative features than the original feature extraction network and alleviates lane-occlusion problems;(2)A feature integration technique is proposed.The technique associates both high-level and low-level information then builds rich-feature vectors via attention mechanism and slicing operation;(3)A lane detection algorithm is proposed.This algorithm employs anchors to gather feature vectors from the extraction network then do classification and regression through fully-connected networks by introducing attention mechanisms.This paper evaluates the effectiveness of the proposed lane detection algorithm on CULane and Tu Simple.Compared with other advanced detection methods,our method shows both a high efficiency and efficacy.