基于深度学习的车道线检测方法研究毕业论文
2021-11-07 20:44:54
摘 要
近年来,随着无人驾驶、自动驾驶技术以及先进驾驶辅助系统等技术的发展,汽车的安全行驶日益成为人们关注的热点。随之,车道线检测作为其中的重要一环,也逐渐成为众多学者研究的热点。目前车道线检测的算法大多是基于视觉,而在现实场景下,可能由于各种因素的影响使得车道线的精确检测非常具有挑战性。因而,车道线检测对于算法的准确性和鲁棒性提出了很高的要求。
为此,本文在了解了深度学习在图像领域的发展情况,结合相关模型,提出一种新的车道线检测算法,改进的U-Net模型。它能较好的识别车道线特征,有较高的准确率和鲁棒性。同时,后续再模型输出二值图的基础上,利用最小二乘法的多项式拟合方式进行车道线的拟合。本文的主要工作如下:
(1)针对车道线检测的图像需要经过复杂的预处理过程,本文设计了一种端对端的车道线检测模型,其编码-解码式的结构能够使输入直接对应输出。其检测过程采用语义分割的思想,结合现有的卷积神经网络和语义分割模型,设计了一种R-U-Net网络模型。它以ResNet和U-Net网络模型为基础,富有模型U-Net的简洁性,又具有ResNet的较深的网络层数。一系列的设计使得模型具有较高的检测精度和较好的鲁棒性。
(2)为了提取车道线在鸟瞰图上的特征,算法在第一部输出二值图的基础上对图像进行逆透视变换,得到车道线的俯视图。随后在车道线俯视图上利用最小二乘法拟合出曲线的相关参数,投影回到原图上即完成了车道线的拟合。
(3)本文提出的算法首先利用二值语义分割网络对原图进行图像分割,然后在利用最小二乘法的多项式拟合对车道线进行提取,从而实现车道线的检测。论文在开源的百度车道线检测数据集上对于算法进行了实验和分析,对其有效性进行了验证。并且在此基础上,更具实际情况调整了相关参数,多模型的相关参数进行了优化,以期达到更好的检测效果。
关键词:深度学习 卷积神经网络 车道线检测 最小二乘法
Abstract
In recent years, with the development of self-driving technology, autonomous driving technology and advanced driving assistance system, the safe driving of automobiles has increasingly become the focus of people's attention. Subsequently, lane detection, as an important part of it, has gradually become the focus of many scholars. At present, most algorithms for lane line detection are based on vision. However, in the real world, the precise detection of lane line may be very challenging due to the influence of various factors. Therefore, lane line detection requires high accuracy and robustness of the algorithm.
Therefore, this paper, after understanding the development of deep learning in the image field, combined with relevant models, proposes a new lane line detection algorithm and an improved U-Net model. It can recognize lane line features well and has high accuracy and robustness. At the same time, based on the model output binary graph, the least square method is used to fit the lane line. The main work of this paper is as follows:
(1) the image of lane line detection needs to go through a complex pre-processing process. In this paper, an end-to-end lane line detection model is designed, and its encoding and decoding structure can make the input directly correspond to the output. The detection process adopts the idea of semantic segmentation, and a r-U-Net network model is designed based on the existing convolutional neural network and semantic segmentation model. It is based on the ResNet and U-Net network models, with the simplicity of model U-Net and the deep network layers of ResNet. A series of designs make the model have higher detection accuracy and better robustness.
(2) in order to extract the characteristics of the lane line in the aerial view, the algorithm performs inverse perspective transformation on the image based on the first output binary map to obtain the top view of the lane line. Then, the relevant parameters of the curve are fitted using the least square method on the top view of the lane line, and the lane line fitting is completed when the curve is projected back to the original figure.
(3) the algorithm proposed in this paper firstly uses the binary semantic segmentation network to segment the original image, and then extracts the lane lines by using the least square polynomial fitting, so as to realize the detection of lane lines. This paper carries on the experiment and analysis to the algorithm on the open source Baidu lane line detection data set, and verifies its effectiveness. On this basis, the relevant parameters are adjusted according to the actual situation, and the relevant parameters of the multi-model are optimized in order to achieve a better detection effect.
Keywords: deep learning; convolutional neural network; lane line detection; least-squares method
目录
第1章 绪论 1
1.1研究背景及意义 1
1.2国内外研究现状 1
1.2.1基于特征的车道线检测算法: 2
1.2.2基于模型的车道线检测算法: 2
1.2.3基于深度学习的车道线检测算法: 2
1.3论文研究内容 3
1.4论文的主要结构 4
第2章 相关理论及技术 5
深度学习概述 5
2.1深度神经网络基础 6
2.1.1单个神经元 6
2.1.2激活函数 7
2.1.3前馈神经网络 9
2.1.4正向传播与反向传播 11
2.1.5损失函数 11
2.1.5梯度下降与其他优化算法 12
2.2.1网络层级结构 14
2.2.2Dropout 17
2.2.3卷积神经网络发展现状 17
2.3语义分割 18
2.3.1FCN 19
2.3.2U-Net 20
2.3.3Mask R-CNN 22
第3章 基于改进语义分割模型U-Net的车道线检测 24
3.1基于resnets残差模块的U-Net网络模型 24
3.1.1特征提取网络 25
3.1.2上采样 26
3.2数据集处理 28
3.2.1数据集的选取和标注 28
3.2.2输入图像的压缩与裁剪 28
3.2.3数据清洗和数据增强 29
3.3算法实验测试 29
3.4车道线二值语义分割 30
3.4.1检测流程 30
3.4.2模型相关参数及训练细节 30
3.4.3模型评价指标 31
3.5实验结果 32
3.6本章小结 35
第4章 基于最小二乘法的车道线拟合 36
4.1逆透视及车道线特征的提取 36
4.1.1逆透视变换 36
4.1.2车道线特征提取 38
4.2最小二乘法 39
4.2.1原理 39
4.2.2最小二乘法的代数法求解 40
4.2.3最小二乘法的矩阵法求解 41
4.3检测效果 42
4.4本章小结 43
第5章 结论与展望 44
5.1结论与收获 44
5.2后期工作展望 44
参考文献 46
致谢 48
第1章 绪论
1.1研究背景及意义
现今,交通运输行业正随着工业化和城市化的发展而飞速发展。道路交通的发展为人们日常生活中的出行提供了极大的便利,有效改善了人们的生活质量,但随之产生的结果就是交通事故的频发,汽车驾驶的安全问题目前已成为全球性的关注主题。相关资料显示,全国的机动车保有量的3.4亿中汽车的占有量达2.5亿。据悉,众多交通事故中九成以上的事故是人为因素所引发的,而自动驾驶能够实时的辅助驾驶员驾驶,甚至特定工况下能够接管驾驶任务,实现无人驾驶。因而自动驾驶可以在一定程度上减少由于驾驶员疏忽大意、疲劳驾驶等因素所导致的交通事故。
由于自动驾驶在缓解驾驶安全方面具有突出的潜力和贡献,所以其日益受到人们的关注。而环境感知作为自动驾驶的重要组成部分,它利用采集到的车内和车外环境,依据相应的算法感知车辆周围的环境,而车道线检测又是环境感知中的重要一环。因此,高级驾驶员辅助系统(Advanced Driver Assistant System,ADAS)系统在这样的背景下应运而生,该系统旨在行驶过程利用传感器感知车辆周围环境,结合GPS、地图等信息进行智能化运算与分析,提前预知可能发生的危险并对驾驶员发出警告,从而提升汽车驾驶的安全性。其中,ADAS的车道保持系统和车道偏移报警系统以车道线检测为基础。