Research On The Lane Detection Algorithm Based On Deep Learning

In recent years,with the rapid development of computer vision and automation technology,autonomous driving technology has gradually entered the public’s vision.Safety is the primary concern of autonomous driving,and the level of safety is inseparable from the ability of environment perception of autonomous driving system,and lane detection is an important link of environment perception.The better performance of lane detection is the premise of the correct planning and decision-making of the automatic driving system,and it is also an important guarantee of the stability and safety of the automatic driving.However,traditional lane detection methods need to be improved in terms of detection accuracy and robustness,and existing deep learning methods also need to balance detection accuracy and timeliness.To solve the above problems,this paper proposes a lane detection method based on deep learning.Specifically,an end-to-end lane detection framework is designed to achieve a balance between accuracy and timeliness.At the same time,a flexible perspective transformation network based on adaptive curve fitting is designed.In addition,lightweight network structure and modules are adopted in the lane line detection network to fully consider the timeliness requirements while improving the detection accuracy.The main work of this paper is summarized as follows:(1)In order to simplify the complexity of lane detection work and avoid the impact of complex curve fitting and other post-processing work on the timeliness of lane detection task,a lane line detection algorithm framework based on end-to-end structure was proposed.In this method,two networks are designed,which are respectively responsible for the realization of lane line instance segmentation and the prediction of perspective transformation matrix under specific lane line scene.The two networks are combined for the end-to-end realization of lane line detection,which effectively solves the problem of the balance between detection accuracy and timeliness.Which are used to implement the lane line instance integral semantic network with lightweight bilateral fusion segmentation as the backbone,the network with the method of integration of deep and shallow network,which has excellent performance in accuracy and timeliness,at the same time,the method by splitting test instance branch differentiate between each lane line,have stronger applicability,the lane detection method can reach 97.72% of the Acc.(2)In order to obtain more accurate lane detection results,a perspective transformation method based on adaptive curve fitting is proposed starting from the network of perspective transformation matrix of forecast image.According to the different shape characteristics of the lane lines,this method adopts an adaptive fitting idea to classify the lane with different bending degrees,and uses different mathematical models to fit them.This method has strong adaptability to the lane lines with different shapes.(3)In order to make full use of the shape characteristics of lane lines,a lane detection method based on hybrid pooling module is designed.The pooling module adopted in this method combines horizontal and vertical strip pooling with the idea of attention mechanism,which can make full use of the remote dependence between pixels in the scene.The design of hybrid pooling module combines the ideas of pyramid structure and strip pooling module and makes full use of global and local information.At the same time,the application strategy of the hybrid pooling module is set up.Different number of hybrid pooling modules are used in the key positions of the lane detection network,and the optimal strategy is determined through experiments,which can bring 0.73% ACC improvement.