Research For Lane Detection And Classification Driven By Visual Attention And Deep Learning

Lane detection and classification is an important module of environmental perception system in automatic driving technology.Traditional lane detection and classification methods generally obtain the location and category of lane by clustering or segmentation methods based on the color,texture and other features of lane.However,traditional methods rely on manually extracted features,and the extracted features can not be applied to lane detection and classification in a variety of road scenes.There are some limitations to these approaches.At present,the approaches based on deep learning regards lane detection task and classification task as semantic segmentation task.However,the main semantic segmentation approaches achieve high performance by stacking a large number of convolution layers,resulting in a large number of parameters and low real-time performance.Due to the limitation of computing resources and storage space of vehicle-mounted devices,the approach is required to have high realtime performance and less parameters.Therefore,in order to achieve higher detection and classification accuracy and real-time with less parameters,the network structure of deep learning is studied and improved in this dissertation.The research contents and innovations are summarized as follows:(1)A fast lane detection network based on mixed attention(Lane-VGG)is proposed.First,the correlation between convolutional features and attention features is established in the encoder to enhance the ability of the network to learn contextual information;second,the contextual information is fused into the upsampling of the decoder to compensate the detail information lost during the information transfer process by the pooling operation;after that,the decoder is divided into two branches,namely the binary segmentation branch that can segment pixels into lane or background,and the spatial structure can be embeded by embeddable branch;finally,the eventual detection map can be gain by fusing the results of two branches.Experimental results show that Lane-VGG can enhance the ability of the network to focus on the lane area by using visual attention,achieve good real-time performance and effectively reduce the false detection rate.(2)A lightweight lane detection network based on semantic segmentation(LLNet)is proposed.First,a simple sub-unit is constructed,which can significantly reduce the amount of parameters of the network;second,in order to enhance the supervision of labels,two sub-units are cascaded to form a dense block using dense connection and skipping structure;and then,the features of dense block from encoder are fused with these from decoder to enhance the network performance;finally,in order to obtain more accurate detection map,the instance segmentation branch is used to replace the binary segmentation branch in Lane-VGG by using the high constraints of lane instance segmentation.Experimental results show that LLNet can effectively improve the detection accuracy and reduce the amount of network parameters while ensuring the real-time performance.(3)A real-time lane classification network based on cascade conception(DCNet)is proposed.DCNet consists of a detection network(DNet)and a classification network(CNet).First,in DNet,the low-level features of the specific position from the encoder are combined with the high-level features from the decoder to realize the joint encoding of high-level and low-level semantic information,and the pixel coordinate values of the lanes are gained;second,according to the pixel coordinate value of the lane,the pixel value of the same position in the image is indexed,and the new feature containing color and texture information is generated;and then,the new feature is associated with the lane type to construct feature-type pairs;finally,the feature-type pair is sent into CNet for training,so as to get the category of lane.Experimental results show that DCNet can achieve high classification accuracy and robustness in a variety of scenes.(4)A lane detection and classification network based on two-directional separated attention(TSANet)is proposed.A smaller convolution kernel is used to construct a symmetric unit,which further reduces the amount of parameters of the sub-units;second,a dense connection block is constructed between multiple symmetric units to obtain multi-scale information;and then,the two-directional separation attention(TSA)is constructed to extract the attention features in both horizontal and vertical directions to establish the dependency between long distance pixels;finally,TSA is embedded into the encoder,and semantic features extracted by TSA are fused into the decoder to obtain accurate lane information.Experimental results show that TSANet can be applied to lane detection and classification in driving scene,and can also realize multi-target segmentation in road scene,which has a good effect of detection and segmentation.