Deep Learning Based Lane Perception And Accurate Position Estimation Method

With the rapid development of science and technology,the automatic driving has gradually become a reality.Regardless of the traditional assisted driving or the current automatic driving,the precondition is that the vehicle can use the sensors to sense the complex environment,so the acquisition of lane information is an important research content.The lane line not only provides the path information for the driving vehicle,but also provides accurate location information.At present,the position information of the lane is mostly obtained by using GPS or curve line fitting based on the lane line perception.The former is susceptible to multi-path effects caused by high-rise buildings,and it is impossible to obtain accurate position information stably.The lane-line sensing algorithms based on traditional machine learning mostly use manual features combined with convolution filtering and Hough transform,which is easily interfered by illumination and other factors,and cannot satisfy multi-scene robustness perception.At present,the lane-line sensing algorithms based on deep learning are mostly computationally complex and cannot meet the real-time requirements of real scenes.Based on the deep learning technology,we propose a real-time deep learning algorithm model,which realizes the end-to-end perception of lane lines.At the same time,in order to improve the vehicle positioning accuracy,the example segmentation technology is incorporated into the deep learning algorithm model,which realizes the initial positioning of the lane line level and proposes a fast lane line accurate position estimation algorithm to realize the lane line level positioning with high precision.The main works are as follows:Firstly,this paper adopts a deep learning segmentation network model based on encoder-decoder architecture to acquire image segmentation perception of driving images,and combines traditional machine learning algorithms to realize segmentation of lane lines.Perceive the results and complete the initial positioning of the lane line level,and which is a preparation for the next precise positioning.In the network model,this paper designs a variety of lightweight methods,reduces network parameters and computational complexity to improve operational efficiency;adopts multiple network performance enhancement strategies to improve model generalization;and uses clustering loss function to implement lane line instance segmentation.Finally,the algorithm is evaluated experimentally.The results show that the algorithm can perform end-to-end lane line sensing in real time on the experimental platform.Secondly,in order to obtain high-precision lane-level position information,this paper transforms the lane line position estimation into a deep learning and easy-to-handle classification task,and then designs a deep neural network to directly estimate the lane line position.This method is an end-to-end solution without relying on cumbersome pre-processing,post-processing etc.Experiments show that the network can achieve sub-centimeter accuracy estimation of lane line position at an average speed of 200 frames per second on the NVIDIA-1080 platform.Finally,In view of the deep neural network for lane detection based on forward camera and lane position estimation based on circle camera,the data sets are expanded and established respectively,which provides effective data support for the acquisition of lane information based on deep learning.