Lane Line Detection Based On Deep Learning Fusion Of Semantic And Geometric Feature Information

Lane line detection is the basis of an advanced assisted driving system such as lane departure warning,which can help determine the vehicle’s position on the road in autonomous driving and provide a basis for trajectory planning.Lane line detection is mainly achieved through traditional feature detection methods and deep learning methods.Traditional lane line feature detection is based on color or ridge line,which can be classified as gradient or variant features,but this method requires manual calibration,cannot be applied to various road conditions,and has poor robustness.With the application of deep learning in the field of computer vision and showing strong capabilities,this technology is increasingly applied to lane line detection.However,because the lane lines occupy a small proportion of the road surface and are easily blocked by vehicles,it is difficult for most models based on deep learning to achieve good results when the marking information itself is insufficient.The purpose of this paper is to build a deep learning model that integrates image semantic information and geometric features to achieve effective lane line prediction.In order to alleviate the problem caused by the lack of information of lane lines in the image,the model in this paper will take the detection of the drivable area of the image as the starting point,predict the number of lane lines by using the geometric information of the lane,and then fuse the two parts of the information to achieve Prediction of lane line location.The main research contents of this paper are as follows:(1)A training set for lane line prediction is constructed.On the open source data set CULane,the car camera images under typical road conditions are selected,including good and insufficient lighting conditions.The labels required for model training are made: including the segmentation map of the drivable area,the number of lane lines,and the marking of each lane line.(2)A model that can fuse semantic and geometric information in images is established.The coding of the semantic information acquisition module is divided into two stages,and LSTM is used to improve the effect of encoding and decoding;a small network is trained,which improves the robustness of perspective changes under various road conditions;Improve the computational efficiency of the model;use the loss function and training method designed in this paper to improve the training convergence speed.(3)A training set was constructed to train and test the fusion model,and the model was verified and optimized.The research shows that: taking the coordinate points of the drivable area as the input,and using the fusion model established in this paper to obtain the fitted parameters of the "coordinate points",the average pixel distance error between the actual abscissa and the predicted and coordinate is within 10 pixels;The segmentation module uses the LSTM secondary encoding method to segment the boundary of the lane area with obstacles more accurately.The fusion model established in this paper can more effectively carry out lane line prediction.