| Nowadays,China is developing faster and faster,all-out poverty has been achieved,people’s living standards are getting better and better,and economic conditions are becoming more and more developed.Thousands of families have their own cars.But with the increasing number of cars,there are more and more traffic accidents caused by improper driving.Therefore,it is necessary to find a safer way for drivers to drive cars.Based on this situation,the autonomous driving ADAS(Advanced Driving Assistance System)technology came into being.As an important direction for the implementation of artificial intelligence,autonomous driving technology has received the attention of a large number of scientific research workers,scientific research institutions and universities.Lane line detection is an important part of automatic driving,and its detection accuracy and speed directly determine the success or failure of automatic driving technology.In the case of good weather,normal road conditions,and non-severe weather with sufficient light,the existing lane line detection algorithms based on traditional and deep learning have good results.However,in the case of high real-time demand,it needs to deal with a variety of road conditions and bad weather such as rain and snow,the existing lane detection algorithms are not satisfactory.How to recognize and detect the lane line in real time under the bad weather and road conditions is a big problem in the field of automatic driving.Based on this topic,this thesis proposes a lane detection model algorithm which can meet the real-time performance and obtain high accuracy.However,in the face of high real-time requirements and the need to respond to various complex road conditions and bad weather such as rain and snow,the existing lane line detection algorithms are not satisfactory.How to detect and recognize lane lines in real time under severe weather and road conditions is a major problem in the field of automatic driving.Based on this background,this paper proposes a lane line detection method that can meet the needs of real-time detection while obtaining high-precision results.This paper analyzes the characteristics of the existing traditional lane line detection algorithms that are poor in robustness and applicability,and the deep learning-based lane line detection algorithm requires high computing power,and demonstrates the necessity and necessity of the lightweight lane line detection algorithm designed in this thesis.Significance,and the reason why instance segmentation is used to detect the lane line because of the narrow and distorted characteristics of the lane line itself.First of all,in order to solve the problem that the existing open source lane line data sets are small,and most of them are for foreign roads,traffic rules and road conditions that are not fully applicable to our Country,manually collected about30,000 data sets including Shenzhen,Beijing,etc.The lane line data set of special road conditions in our Country,and based on the situation that the lane line detection algorithm in this paper needs to be deployed in the low computing power on-board processor,a set of lightweight lane line data set labeling rules is designed.And through the normalization of the input image,special ROI(Region of Interest)design,scaling,Gaussian blur and other technologies,the traditional deep learning preprocessing part is improved,and the data preprocessing part is further enhanced.And compared with the open source Tucson data set,according to the experiment,the lane line data set in this thesis is more suitable for the lane line detection environment in my Country than the foreign open source data set.Secondly,in order to solve the problem that the existing deep learning-based lane line detection algorithm requires too much computing power,this paper has carried out a series of optimization and improvement on the basis of Lanenet so that it can meet the requirements of the on-board processor.Requirements for real-time lane line detection.The improvements are as follows:(1)The innovation of the basic backbone network,drawing on the ideas of Mobilenet and yolov4,redesigning the detection backbone network,replacing the original Enet,and improving the characteristics of the model while reducing the amount of calculation Extraction ability;(2)Use the LMish activation function as the activation function of the network to improve the nonlinear performance of the model;(3)Use a new learning rate adjustment strategy to train the network model,so that the model training process converges faster and the detection effect is better;(4)Accelerate and compress the detection part of the model.In order to be deployed on the car and machine side,this paper combines the convolutional layer and the batch standardization layer,as well as various post-processing methods,so that the algorithm in this paper is detecting The amount of calculation required for lane lines is reduced by one third;(5)The use of Npz format files to train the network increases the training speed by 60%.After the above innovations and optimizations,the lane line detection algorithm in this thesis has a greater lead in various indicators than the existing lane line detection algorithm.Finally,the lane line detection points are clustered by Mean Shift clustering algorithm based on lightweight design,and the light-weight post-processing algorithm is used to fit them.Through experimental analysis,it can be seen that the lane line detection algorithm in this thesis is nearly 40 times faster than FCN-8,nearly 20 times faster than VPGNet,and nearly 10 times faster than Lanenet,the lane line detection algorithm in 2018,even if it is Compared with Lanenet,which uses Google’s Mobilenet,it is still nearly 7 times faster.When using high-end car processors such as Kirin 990,the average detection time of a single frame is only6.42 ms. |
PDF Full Text Request