Design And Research Of Line Detection Algorithm In Rainy Days Based On Deep Learning

Lane detection is one of the most important tasks in autonomous driving system,and the accuracy,real-time and stability of lane detection are crucial to the overall performance of the system.In recent years,deep learning,with its great fitting and generalisation capabilities,has driven the development of lane detection methods,which have better accuracy and robustness than traditional lane detection methods.In practice,the vehicle driving scenario is complex and variable,especially in bad rainy weather conditions lane detection is still difficult,because rain streaks will make the images collected by the autonomous driving vision system appear lane blurring,lane occlusion and lane visibility decline,resulting in the image to be detected showing significant quality degradation,seriously affecting the accuracy and stability of the lane detection task.This paper designs an algorithm for lane detection in rainy weather scenarios,which will help to solve the problems faced by lane detection in rainy weather conditions.The main research work in this paper is as follows.(1)A backbone network for the lane detection algorithm is designed based on a lightweight ENet segmentation network,which can output lane semantic segmentation graphs and instance segmentation graphs,and a pyramid pooling module and a soft attention mechanism are introduced to make various improvements to the backbone network to obtain several lane detection networks with different characteristics.In terms of dataset processing,data augmentation is used to augment the dataset,while the Focal Loss function is introduced as the semantic segmentation loss function in order to address the problem of imbalanced lane sample distribution.Finally,several sets of comparison experiments were conducted,and the results showed that the ENet+PPM+SE network designed in this paper is a better comprehensive network,with segmentation accuracy indicators m Iou as high as 79.3%,F1 as high as 74.6%,and processing frame rate of about 45 FPS.(2)Combined with the depth separable convolution,the pixel by pixel filtering rain removal algorithm is lightweight improved.The main improvement point is to replace the standard convolution of the KPN network included in the algorithm with depth-separable convolution,thus obtaining the improved network(KPN＿Tiny).In terms of dataset,the Rain Lane9000 dataset containing nine rain patterns was synthesised based on the Tusimple lane dataset,which can be used for training and testing of the rain removal network.The experimental results show that the improved KPN＿Tiny network reduces the number of parameters by about 88% and the number of operations by 76% compared with the KPN network,and the processing frame rate is about 27 FPS;in terms of rain removal quality,the KPN＿Tiny network performs almost the same as the KPN network,and can better improve the image quality degraded by rain patterns.(3)A lane detection algorithm including the rain removal link of lightweight image is designed.The experimental results show that the lane detection algorithm designed in this paper improves the lane detection accuracy index m Iou by 5.0%,F1 by 9.3% and the processing frame rate by about 21 FPS in various rain test environments compared with the lane detection algorithm without rain removal links.