Research On Lane Detection Method Based On Deep Learning

As a key component of automatic driving system,Lane detection plays an important role in real-time vehicle positioning,driving path planning,lane keeping assistance and adaptive cruise control.Due to the inherent elongated characteristics of lane,it is easy to be affected by complex environment such as severe occlusion,bad weather conditions and blurry road surface.This thesis proposes two different methods for lane detection in complex scenes.(1)A lane detection method based on spatial feature interaction was proposed,named SFINet.SFINet network consists of main branch and auxiliary branch.The main branch corresponds to the classification task based on row direction,and the auxiliary branch corresponds to the segmentation task.In order to better deal with complex scenes,the main work of SFINet is as follows: In view of the problem that simple features extracted by Res Net cannot cope with complex scenes,the spatial feature interaction module is added after Res Net to capture the spatial relations of pixels between rows and columns,so as to improve the feature extraction ability of network.In order to solve the problem of rough upsampling result of ordinary bilinear interpolation in segmentation task,a bilinear interpolation method combined with transpose convolution is introduced to make the upsampling result more refined.Aiming at the uneven distribution of lane line and background pixel samples,Focal Loss was used to increase the weight of difficult samples and make the model pay more attention to the lane line pixels that are difficult to detect.The experimental results show that the SFINet method can effectively deal with lane detection in complex scenarios,and the F1 score detected on CULane dataset reaches 73.8%.(2)A lane detection method based on dual attention mechanism was proposed,named FDANet.In order to give consideration to the accuracy and speed of detection,the spatial feature interaction module with large computation was removed from SFINet,and the dual attention mechanism was integrated to detect lane lines.The dual attention mechanism models feature correlation in spatial dimension and channel dimension,which can improve the network’s feature extraction ability of lane lines in complex scenes,so as to better deal with the detection of complex scenes.In addition,the double attention mechanism is introduced,which only slightly increases the amount of computation and can ensure the real-time detection.The experimental results show that FDANet method can not only effectively deal with complex scenes,but also faster,providing a reference scheme with both accuracy and speed for practical application.