Research On Lane Detection Algorithm Based On End-to-end Convolution Neural Network

In the field of unmanned driving,identifying lane lines from real driving scenes is an important part of autonomous driving vehicles understanding the surrounding driving scenes.The quality of the identified lane lines has a huge impact on the division of the vehicle’s drivable area and driving decisions.In order to ensure the driving safety of unmanned vehicles,unmanned vehicles need to have the ability to drive normally in different types of driving scenarios(crowded,night,dazzle lighted scenes,etc.).At the same time,the surrounding environment may also change uncontrollably at any time when the vehicle is traveling in spatial,and factors such as nearby pedestrians and vehicles need to be considered by the algorithm.Therefore,the type of driving scenes and the dynamic changes in the driving scene during the driving of the vehicle will bring great trouble to the lane detection algorithm.Moreover,most existing lane detection models cannot clearly detect unclear and occluded lane lines,so this thesis proposes to use the end-to-end convolutional neural network to detect lane lines in complex driving scenes by semantic segmentation.The main work is as follows :(1)A lane detection algorithm based on attention mechanism is proposed.Inspired by the spatial convolution block,the proposed model optimizes the process of information transmission in the spatial.Using the attention mechanism after spatial convolution can not only increase the information exchange between pixels in the channel dimension,but also weight the extracted important features.By combining the spatial convolution with the attention mechanism,the performance of the model is improved.(2)A robust end-to-end lane detection model using vertical spatial features and contextual information in the complex driving scenes is proposed.In order to make more efficient use of the contextual information and vertical spatial features in the driving scenes,the feature merge block and information exchange block are designed to the make the model better detect unclear and occluded lane lines,so that the model can be used in complex driving scenes more robust.(3)Comparison experiments with mainstream models are conducted on two lane detection datasets,which are the CULane dataset and Tu Simple lane detection benchmark dataset.And the experimental results were evaluated in detail qualitatively and quantitatively.Through the analysis of the experimental results,the model proposed in this thesis performs better in driving scenes.