Research On Measurable Continuous Scene Lane Line Extraction Based On Deep Learning

In recent years,the improvement of living standards has led to a sharp increase in the number of cars.The frequent occurrence of various traffic accidents is following by it.Auto-driving can reduce traffic accidents caused by manual errors and save a lot of manpower,which is one of the popular research directions in the automotive field.An important aspect of automatic driving is Lane line detection.Lane line detection can standardize the driving route,give emergency instructions on the deviated driving route,and avoid traffic accidents caused by fatigue driving or wrong driving.The experimental data of lane detection in this paper are some street scenery images collected from MMS mobile road survey in Beijing.In this paper,the traditional lane detection algorithm and the lane detection algorithm based on deep learning are experimented,and the experimental results are compared.The main contents are as follows.The traditional lane line detection algorithm based on artificial design mainly combines threshold segmentation of mathematical morphology and Hough transform to detect lane line.Threshold segmentation separates lane lines from background.Mathematical morphology can connect Lane line breakpoints and remove noise points.Hough transform detects lane lines by matching points in parameter space with lines in image space.Lane detection algorithm based on deep learning is the core of this paper.The U-net network and VGG16 network are fused mainly through migration learning,which changes the input and output of the network,as well as some hidden layers in the middle,so that the network is more suitable for the data in this paper..It mainly includes Annotation of Data Set,Design and optimization of network structure,network training and so on.The labeling of data sets is crucial to the accuracy of training results.In the experiment,the labelme software was used to label the training data set manually.The network structure mainly combines the network structure of U-net network and the data extraction part of VGG16 network.VGG16 has a good effect on feature extraction,and U-net network can combine advanced features with low-level features to output more perfect features.Combining the advantages of the two networks and according to the characteristics of this experiment,the experimental network is obtained by fine-tuning.The experiment optimizes the network throughAdam algorithm,and gets the lane detection network.This paper divides the data into five different data sets according to the quality of lanes,and makes a comprehensive comparison between the two algorithms according to these data sets.With the image label as the standard,the traditional lane detection is accurate when the lane location is correct and the number is more than 80% of the number of lanes labeled.The deep learning algorithm calculates the value of IOU.If the value of IOU is more than 0.8,it is regarded as the detection accuracy.The overall detection accuracy of the traditional lane line detection algorithm based on artificial design and Hough transform is 52%.The detection results are poor.There are many missed and mistaken detection,and the running speed is slow.Unlike the former,the detection accuracy of the lane detection algorithm based on deep learning can reach90%.It has better effect and stronger applicability for various data sets,and its running speed is nearly 20 times faster than the manual design algorithm,which almost meets the real-time requirements.