Research On Lane-line Detection Method Based On Convolutional Neural Networks

As one of the key technologies in the environmental perception system of intelligent vehicle,lane-line detection technology has high study value.The traditional lane-line detection method cann't already satisfy the precise demand under complex environment,so in this thesis,convolutional neural network is used to deeply study the lane-line detection technology.Firstly,an ENet-based lane-line segmentation method is proposed.After studying the theoretical basis of convolutional neural networks,several classical semantic segmentation networks are compared.From the specific needs of lane-line detection,the lightweight ENet is selected as the basic network for lane-line segmentation,and the initial ENet model is completed.Compared with other semantic segmentation networks,ENet has a simpler network structure and better real-time performance,so it is more suitable for lane-line segmentation task.Secondly,a lane-line detection method based on S-ENet and least squares fitting is proposed.The common problems of several difficult scenarios for lane-line detection are analyzed.And an improved ENet model,S-ENet,is proposed to enhance the ability to extract spatial context information of image,thereby the model's robustness to complex environments is improved.Then the lane-line segmentation result is converted to a bird's eye view by inverse perspective transformation,and the lane-line fitting is performed by using the sliding window and the least square method.Finally,the superiority of the lane-line detection algorithm in this thesis is verified by experiments.After determining the lane-line detection dataset and evaluation criteria,the lane-line segmentation model is trained,then the trained model is tested on the test set.The test results show that the S-ENet model has higher stickiness and better segmentation effect in complex road scenes.At the same time,the contrast results between the proposed model and the traditional threshold segmentation model show that the convolutional neural network has great advantages in detection accuracy and speed.In addition,by constructing a real vehicle test platform to collect campus driving video,the generalization and practicability of the lane-line detection algorithms are evaluated.The experimental results show that the lane-line detection method based on S-ENet and least squares fitting has better performance,can process a frame within 69 millisecond with an accuracy of 92.8%.In this thesis,the lane-line detection algorithm based on convolutional neural network has strong adaptability to complex road scenes,has high detection accuracy and basically meets real-time requirements,and has good theoretical significance and practical value.