Research On Multi-layer Lane Detection Method Based On Residual Attention

In recent years,the field of unmanned driving has received more and more attention,and the task of lane line detection as its core task has also become a hot field of scientific research.The purpose of lane line detection is to detect lane lines in complex traffic scenarios and provide decision-making for unmanned driving.The current lane line detection methods mainly have the following two difficulties.On the one hand,the existing lane line detection algorithms are based on the segmentation method to classify the pixels of the lane line image,the calculation amount is too large,and the detection speed performance cannot reach real-time performance.On the other hand,the network model is flawed,leading to the loss of key information in the local receptive field;due to problems such as illumination,blurring of the lane line,and obstacles blocking the lane line,the accuracy of detecting the semantic lane line is not high.Aiming at the above two difficulties,this thesis proposes a multi-layer lane line detection method based on residual attention.The main work is as follows:(1)Aiming at the slow speed caused by the segmentation of lane line pixels in the lane line detection technology,a multi-layer-based lane line detection method is proposed.That is,divide the layers with the same number of lane lines,and classify the only lane line of the layer.This thesis regards the lane line detection task as a classification task of the lane line in the row direction(the row direction refers to the horizontal direction of the content shot by the driving recorder),and divides each row into a fixed number of components,and specifies the layer corresponding to the lane line In the row direction,output the probability that each component has a lane line point,which is represented by a multi-dimensional vector of the behavior unit.The geometric characteristics of the lane line are used to constrain the lane line points of each layer of the lane line to realize the lane line detection.Multiple layers can ensure the comprehensiveness of lane line detection.Compared with the segmentation method for processing lane line image pixels,this method only calculates the multi-dimensional vector of a specific line,which can greatly reduce the amount of calculation.(2)Aiming at the low accuracy rate caused by the loss of key information of the lane line,this thesis adds the residual unit form to the attention mechanism,and proposes a lane line detection method based on the residual attention mechanism.The method includes two branches,the feature processing branch is used to extract the input image features;the residual attention branch uses the hourglass network to up-sample and down-sample the input image to extract the key information weights of the features.Through the two branch point multiplication operations,the influence of noise in the deep network on the update gradient can be reduced,and the robustness of the model can be enhanced.The model uses the entire input image as the receptive field to ensure that the key point information is not lost and improve the accuracy of lane line detection.(3)This algorithm has been experimented on the public data sets Tu Simple and CULane,and the visualized results of the experiment have been analyzed.Through ablation experiments on the residual attention module,multi-layer lane line detection module,and lane line structured loss function,it is proved that each module of the algorithm has an important role in improving the performance of lane line detection;the baseline experiment proves the algorithm Its performance is significantly better than that of lane line detection algorithms such as SCNN and Lane Net.