Lane Recognition In Complex Scene Based On Semantic Segmentation

Lane detection is one of the key tasks of autonomous driving,and also the difficulty in the practice of autonomous driving.In real life,lane lines are irregular,difficult to predict,complex background and so on,so it is a very complicated process to detect them effectively.In order to solve the problem of lane detection in complex scenes,a hybrid architecture based on deep learning is proposed in this thesis.This hybrid architecture can take continuous lane detection scene as input,make full use of time series information in video stream,and implement lane detection by semantic segmentation method.The main tasks completed include:(1)Based on Res Net50 residual network and UNet semantic segmentation codec architecture,Res Net50 is used as the backbone network of UNet encoder to construct Res＿UNet model and optimize the performance of UNet network.This model solves the problem of gradient disappearance and performance degradation caused by network depth and can better combine the semantic information of context to achieve multiscale segmentation.(2)In order to solve the problem of lane detection in complex driving scenarios,a hybrid architecture UNet＿Conv LSTM was constructed on the basis of Res＿UNet network and Conv LSTM,so that it could combine the lane timing information in continuous frames and then accurately detect lane lines in complex environments.(3)A strip pooling suitable for the specific shape of lane lines is introduced,so that the backbone network can effectively model the remote dependency relationship,which helps to capture the local context,avoid the influence of irrelevant areas on label prediction,and realize the accurate segmentation of lane lines.Based on the strip pooling operation,the remote context information is collected from different spatial dimensions by horizontal and vertical strip pooling operation,and the strip pooling module(SPM)is constructed.Next,a mixed pooling module(MPM)is created by combining pool operations of different kernel shapes,which can aggregate different types of context information to make feature representation more differentiated.Finally,the strip pooling operation is introduced into the attention mechanism to create the strip attention module(SAM),which can not only capture the non-local context,but also reduce the computational complexity in time and space.(4)Insert strip pooling related modules into mixed architecture UNet＿Conv LSTM network,test and analyze the influence of strip pooling modules on network performance based on different data sets to determine the selection and insertion location of modules.Finally,the hybrid architecture UNet＿SAM＿Conv LSTM of lane detection algorithm was determined,and good detection results were obtained in complex scenes.The semantic segmentation algorithm of lane segmentation based on hybrid architecture designed in this thesis integrates Res＿UNet based on codec structure and Conv LSTM which can extract time sequence information,and adds related modules of strip pooling which can aggregate context information to complete lane recognition in complex scenes based on semantic segmentation.