Research On Lane Detection Method Based On Deep Learning

With the rapid development of the automobile industry,automatic driving has become a research hotspot in recent years.Lane segmentation and road detection are very important research fields of automatic driving system.Traditional lane segmentation and lane detection methods usually need to be arranged artificially according to scene characteristics,which is difficult to cope with complex and changeable road environment.The extensive development of deep learning methods has accelerated the research process of lane segmentation and road detection.Therefore,this paper studies lane detection method based on deep learning,mainly including lane semantic segmentation and lane detection.Firstly,this paper introduces the basic theory-level related technologies,including forward and backward propagation algorithms,activation functions,convolution and pooling of deep neural networks.The paper also introduces the network platform and experimental environment in order to pave the way for the research content of the following chapters.Secondly,in order to improve the low accuracy of road segmentation in the complex environment of the current convolutional neural network,a lane semantic segmentation network is constructed based on based on the coding and decoding network structure.Aiming at the problem of detail information loss caused by the existence of pooling layer in common convolutional neural networks,a combination of hole convolution and spatial pyramid pooling is adopted to fuse multi-scale feature information while increasing receptive field and not reducing resolution,so as to improve the segmentation performance of the network.In order to select the loss function suitable for lane division tasks,verify the influence of each network module on network performance and prove the superiority of the divided network,comparative experiments are designed from three aspects: loss function,network module and different networks.Finally,aiming at the problem that the current single lane dete ction method based on road segmentation or lane marking has poor effect in multi-lane detection tasks,a multi-task network model is constructed based on the above lane semantic segmentation network.The two branches of the model share encoders,which can realize lane line instance segmentation and road semantic segmentation simultaneously.Then,a training strategy is designed to realize the alternate training of each branch so as to ensure the sufficiency of training of different branches.Secondly,a fusion algorithm is proposed to fuse the results of road segmentation and lane line instance segmentation to obtain the detection results of the center lane and the adjacent left and right lanes.Finally,comparative experiments are designed respectively from three aspects of road segmentation,lane line instance segmentation and multi-scene multi-lane detection to verify the effectiveness of the scheme.