Research On Lane Detection Based On Deep Learning

Lane detection is an important part of the advanced assisted driving system.At present,the current number of cars is increasing,and traffic accidents occur frequently.The advanced assisted driving technology has attracted widespread attention.The research of fast and accurate lane detection technology can improve the accuracy of lane keeping and lane departure warning system,which is of great significance for improving driving safety.Traditional lane detection methods rely on a combination of highly-specialized,hand-crafted features and heuristics,usually followed by post-processing techniques that make it difficult to cope with complex road scenes in real-world situations.Due to the excellent performance of deep learning technology in the field of computer vision.In recent years,deep learning technology has also been applied in the lane detection task,and has achieved better results than the traditional method.At present,the lane detection algorithm based on deep learning technology still faces some difficulties.The appearance characteristics of the lane are not obvious,and it is easy to be blocked by the front car or the interference of shadow and illumination.The characteristics of the local pixel alone are not enough to obtain better results.In view of the above difficulties,this thesis studies the semantic segmentation of lane lines based on convolutional neural networks.The details are as follows:(1)A novel fast spatial attention module is proposed,which convolves on the spatial dimension of the feature map to obtain long-range spatial context information.Through the long-range spatial context information,the situation that the lane line part identification feature is not obvious or invisible is improved.(2)Design a module combining channel attention and classification auxiliary supervision loss to utilize the correlation between channels and predict the existence of lanes of the whole picture.Such supervision can guide the network to understand global semantic information and assist the network to obtain more discriminative features.(3)Using the weighted cross entropy loss function to improve the serious imbalance of the number of lane line pixels and background pixels in the picture,and give a larger weight to the number of pixels with fewer lane lines.Avoid learning from the network biased towards the background.(4)Using the online hard example mining strategy to improve the problem that the number of pixels of the simple sample in the training picture is not equal to the number of pixels in the difficult sample,discarding some of the too simple pixels to make the network focus on the difficult samples' optimization.According to the improved method proposed by us,the experiment is carried out on CULane,the largest open lane dataset,and compared with other advanced methods.The experimental results show that our method is effective and efficient.