Research On Lane Detection And Following Method Of Autonomous Vehicle Based On Deep Learning

Lane line detection and following in driving scenes is an important component of unmanned vehicles and advanced driver assistance systems.With the rapid development of artificial intelligence technology,the combination of unmanned vehicles and deep learning technology has gradually replaced traditional methods,and is widely used in unmanned driving perception,control and decision-making fields,effectively reducing more traffic problems and improving the driving safety of vehicles.In recent years,various complicated lane line detection and following methods have proposed one after another,however,most methods often have poor performance when dealing with extreme difficult situations such as severe shadows,dark nights,and missing markings.The perception,control,and decision-making modules of unmanned vehicles are closely related,and the accuracy of lane line detection seriously affects the accuracy of lane following and directly threatens road traffic safety.Firstly,complete data collection.A method of lane marking based on vanishing point is proposed.For typical road conditions such as night,wireless and shadow,the simulation and real lane detection and lane following data sets are made,which are consistent with Chinese road.Secondly,a multi-frame lane line detection method based on UNET＿CLB is proposed.Aiming at the problem of poor recognition of difficult situations such as wireless,dark night,shadows,etc.,on the basis of traditional deep learning technology,the introduction of multiframe information of continuous driving scenes for lane line detection.In order to solve the problem of high-level semantic information loss in convolutional neural network(CNN),CNN is combined with convolutional long-term memory network convlstm and deep dense connected convolutional network(DENSE＿NET),a deep high-level semantic extraction network is proposed.Experiments on the Tu Simple and CULane datasets show that the detection effect of this method is good in both robustness and recognition accuracy.Thirdly,a spatiotemporal lane following method based on LSTM＿DTS is proposed.Aiming at the problem of the loss of key information in long-term sequences of cyclic neural networks,TIME＿ATTENTION(time attention network)is used to predict the correlation of input information and SPACE＿ATTNENTION(spatial attention network)is used to obtain effective historical information.Introducing vehicle kinematics constraints to improve the feasibility of following control signals.Test results on real data sets show that the method in this paper can accurately simulate human driving behavior(steering wheel angle and vehicle speed),while improving the continuity and stability of control behavior.Fourthly,the simulation verification is carried out based on the Webots simulation software.An unmanned road simulation model is built,and the lane line detection method UNET＿CLB proposed in this paper is combined with the lane following method LSTM＿DTS.At the same time,experiments are performed on simulation and real environment data sets.Simulation and experimental results show that the proposed method can achieve unmanned vehicle lane line detection and lane following accurately under the bad conditions of night,shadow,wireless and so on.