Research On Lane Detection Algorithm Based On Instance Segmentation

With the development of social science,many cars are equipped with Lane assisted driving.Lane assisted driving is like another eye of the driver.It can keep the vehicle in the lane line or give out departure warning to the driver when it deviates from the lane line.In addition,the development of artificial intelligence provides technical support for the landing of autopilot.Lane detection is one of the key technologies in the perception module of autonomous driving vehicle.By identifying the lane line,it provides real-time information basis for decision-making and planning module of autopilot.In order to ensure the lane line can be detected accurately and in real time,the research and improvement of lane line detection algorithm has been developing.The traditional lane line detection system is designed complex and has a large amount of computation.It can be detected by the methods of custom and manual feature extraction.At present,it is a trend to detect lane line by deep learning.It is better than traditional method in calculating speed and accuracy,and can deal with different characteristics of lane line.Generally,the method based on deep learning can only detect a fixed number of lane lines,and can not deal with the changes of different lane lines.In this paper,in depth learning,the lane line is detected by case segmentation method.Each lane line after segmentation forms its own example,which can cope with the changes of different lanes.The method has good performance in the accuracy and speed of lane line detection,and promotes the further development of lane line detection technology.The lane detection algorithm model based on instance segmentation designed in this paper absorbs the characteristics of the current semantic segmentation model structure.Aiming at the distribution characteristics of large space span of lane lines,not only the network needs to have strong global feature extraction ability,but also a good receptive field for the details of the network.Through the use of bilateral segmentation network structure,lane detection can be well satisfied Therefore,the robustness of prediction results can be improved by using this network structure.We design a instance segmentation two branch network model,and use bisenetv2 bilateral instance segmentation network to extract the characteristics.The network can extract the specific feature information of lane line through detail branch and semantic branch.Then,the lane line pixel can be accurately extracted,and then the binary branch and embedded branch of the case segmentation are used to cluster into Lane instances.In the segmentation lane line,it is used to fit the vehicle Before the road,we further designed a selective lane line fitting algorithm.In the actual road,due to the diversity of lane line types,according to the advantages of polynomial fitting,the fitting line is automatically selected to ensure the accuracy of the fitting.The algorithm has been verified in tusimple data set and good results are obtained.The experiment shows that the algorithm proposed in this paper has strong detection effect and robustness.Even if the lane line is blocked,it can be well responded to.At the same time,the algorithm has a fast detection speed.The detection accuracy of the example segmentation network designed in this paper can reach 67.8% Miou(even intersection ratio),and the detection algorithm of the whole instance segmentation lane line can reach 63 FPS detection speed meets the requirements of real-time and accuracy.