Research On Lane Detection Of Intelligent Vehicle Based On Deep Learning

The solution of the environmental perception problem is one of the important prerequisites for the realization of intelligent vehicle automatic driving,and the research of lane line detection occupies a key position in the environmental perception.With the development of deep learning technology at home and abroad,and in order to minimize the impact of environmental factors on lane line detection tasks,researchers began to use this technology for lane line detection research.In this thesis,lane line detection is regarded as an image segmentation problem,and lane line detection is researched based on deep learning algorithms to cope with the challenges of rapid changes in the environment during driving.First,based on BiSeNet V2,a two-branch neural network model is designed for lane line image segmentation.This example segmentation branch network includes two branches,namely semantic segmentation branch and clustering branch,to improve detection ability.The method is to operate on discrete lane line feature points;then,using the fitting method of combining straight lines and curves,for a certain field of view,select the corresponding feature points and adopt effective robust fitting to obtain a suitable.Finally,the model is trained and verified based on Tu Simple and other data sets,and the model is integrated into the ROS(Robot Operating System)platform for the task of lane line detection of intelligent vehicles.At the same time,After improving the probabilistic Hough transform algorithm commonly used in traditional methods,the algorithm based on the BiSeNet V2 neural network model is compared with it on the smart vehicle.The accuracy of the algorithm is about 3.9 times that of the improved probabilistic Hough transform algorithm.The detection speed is about 2.9 times.All in all,this thesis proposes a real-time lane line detection algorithm based on deep learning that can be applied to smart vehicles.The experimental results show that the lateral offset of the algorithm based on the BiSeNet V2 neural network model is only about 2.5 pixels,and its average intersection ratio is up to 72.6%,which can be used for smart vehicles to respond to lane changes during driving.