Research And Implementation Of Lane Detection Algorithm Based On Improved Convolutional Neural Network

In recent years,with the increasing improvement of people’s living standards and the rapid development of China’s automobile industry,there are some increasingly serious traffic problems have been brought by the sharp rise in the number of vehicle ownership.Road safety has become an important social issue that is of concern to a wider range of people because of its broad reach.The rapid development of artificial intelligence technology makes autonomous driving possible,and the Advanced Driving Assistance System of vehicles has become a hot direction of research nowadays.As one of the core functions of Advanced Driving Assistance System,lane detection technology can detect the road situation in real time during the driving process,and when the vehicle deviates from the current lane,it can issue a timely warning to remind the driver or let the automatic driving vehicle make timely operation,which can effectively reduce the occurrence of traffic accidents and thus more effectively protect the driver’s life health and personal safety.Therefore,the study of lane detection is of great significance to improve the safety of driving.In actual traffic road scenarios,lane detection is challenging due to the complex environments,weather changes,lighting effects,and broken and obscured lane lines.Existing lane detection methods focus less on the simultaneous detection of multiple lanes and ignore the differentiation of lane types.At the same time,there are also have disadvantages such as low detection accuracy,single detection scenario and poor real-time performance.In order to solve the above problems,This thesis proposes an improved Deep Lab v3+ based lane detection algorithm,which achieves end-to-end lane semantic segmentation.The main work and innovation points of This thesis are as follows:(1)To address the problem that lane lines account for a relatively small number of pixels in road images and often present a thin and narrow shape,the Deep Lab v3+ encoder part has been improves to effectively extract the semantic features from images.In the encoder,the modified Res Net is used as the feature extraction network firstly.The modified Res Net improves on the Res Net-50 by deepening the number of network layers for better lane line feature extraction,and incorporates the channel attention mechanism and the atrous convolution.Using the modified Res Net with deep layers as the feature extraction network,more valuable deep-level semantic features can be effectively extracted;In addition,the Atrous Spatial Pyramid Pooling module is lightweighted and the depthwise separable convolution is used to replace the original ordinary convolution,which can effectively reduces the parameters of the overall model and ensures the real-time performance of the model.(2)The feature fusion module in the original Deep Lab v3+ decoder is improved,and a decoder incorporating a multilayer low-level feature fusion is designed.Only one feature fusion was performed in the original Deep Lab v3+ network,and large multiplicative upsampling was directly used to restore the feature map to the input size.However,such an operation is prone to feature loss and loss of lane line edge detail information,which affects the final detection results.The improved multilayer low-level feature fusion module uses small multiplicity upsampling to gradually recover the feature map size and performs multiple feature fusion operations with the low-level feature map output by the feature extraction network,which can effectively fuse feature information at different scales and obtain more low-level detail information while retaining high-level semantic information.(3)Since the initial labels classified by the dataset used in This thesis have 33 classes,This thesis uses recoding to divide the labels into 8 classes which includes background and 7lane line types,in order to facilitate the detection segmentation of the network model.In addition,a variety of data enhancement methods are proposed,including pixel-level enhancement,graphical enhancement,shape enhancement and random cropping operations,which can effectively improve the training effect of the network model and make the detection results more accurate by using these data enhancement methods before sending the images to the network model for training.(4)According to the practical application requirements,a lane detection system with a visual interface is designed and implemented in This thesis.Users are able to select the images need to be detected and upload them into the system,which first automatically loads the trained network model,then pre-processes the images before send them to the model,and displays the segmentation results of lane detection finally,thus facilitating users to conduct application research on this basis.Finally,This thesis carries out a variety of comparison experiments as well as ablation experiments on the actual road image dataset to comprehensively evaluate the effect of the improved modules,which proves the feasibility and superiority of the algorithm through experiments,and has some practical application value.