Research And Application Of Lane Detection Method Based On Deep Learning

Lane detection is one of the crucial tasks in automated and assisted driving systems.Lane lines are important signs to guide the vehicle and help maintain a stable trajectory,which can reduce traffic accidents caused by lane deviations.Therefore,the research on lane line detection is of great practical value.However,realistic scenarios where lane lines are obscured,broken,spatially spanned,and disturbed by lighting conditions pose a challenge for lane line detection.At the same time,the lane line detection task has high real-time requirements.To address these issues,this thesis investigates lane line detection methods from two mainstream ideas based on line classification and semantic segmentation.The work in this thesis consists of the following three parts:(1)For the real-time requirements of the lane line detection task and the problem of novisual cues,a fused attention module-based method for line classification lane line detection is proposed.In the main branch,a lightweight residual network is first used as a feature extraction network to reduce the computational effort,and deformable convolution is introduced in the residual blocks of some of its layers.It makes the shape of the convolutional kernel and lane lines more compatible,thereby obtaining the refined structure of the lane lines and improving the ability which about feature extraction of the model.Secondly,the horizontal symmetry feature of lane lines is modeled through the feature flipping fusion module.In addition,the efficient channel attention and the improved spatial attention module are designed to be tandem to form the fused attention module.The purpose is that it can guide the network to obtain spatial contextual information to understand global semantic information and effectively solve the problems of lane line defects and occlusions.Finally,a fast detection method based on line classification is used to output the detection results.In addition,the auxiliary branch goes through the spatial attention module to capture the long distance information association in space,performing pixel-by-pixel semantic segmentation of the images,and the generated results are used to assist the network training.Displaying images through experimental effects can intuitively verify the improvement in speed and accuracy on two different datasets in this thesis in two different datasets.(2)In order to solve the problem that the output dimension of a general semantic segmentation network is equal to the number of pre-defined segmentation target classes which cannot cope with the changing number of lane lines on different roads.Conduct in-depth research on the semantic segmentation idea based on auxiliary branches used in the above methods,and a lane line instance segmentation detection method based on spatial feature aggregation is proposed.Firstly,the method uses feature pyramids to perform multi-scale feature fusion on the extracted features and obtain local details and high-level semantic information.At the same time,a multiscale decomposition convolution is designed to expand the network perceptual field so that the network can better extract lane line features.Secondly,the spatial relationship of pixels across ranks is captured using a spatial feature aggregation module,and the features in both dimensions are fused and enhanced.Finally,a two-branch decoding network is used to achieve multiple lane line segmentation decoding and clustering of pixel points to complete instance segmentation of different lane lines.The method shows good segmentation results on the large multi-complex scene dataset which is the CULane.(3)Based on the above research methods,a lane detection system was developed using PyCharm 2021.3 and the PyQt5 framework,which integrates the two methods mentioned above for the users to choose and complete the detection for both pictures and videos.