基于深度学习的夜间前方车辆识别研究毕业论文
2021-03-29 22:36:49
摘 要
随着社会的发展与进步,我国的汽车保有量稳步提升。汽车数量的增加也产生了交通阻塞、环境污染、交通事故频发等一系列的社会问题。特别是在夜间,交通事故的发生率远远高于白天,为了提升交通安全系数,增强汽车安全性能,夜间前方车辆识别技术的研究对于自动驾驶系统、自动识别技术的发展都具有重要的意义。
本文主要是基于前向单目相机对前方车辆进行图像数据提取,利用基于深度学习技术进行夜间前方车辆的识别研究。论文首先对国内外车辆识别技术的研究现状,以及车辆识别技术的发展进行综合阐述,认为夜间光线较弱,车辆识别率受环境、车速、设备影响较大,因此,论文选取尾灯提取方法来识别车辆,首先对所提取的车辆图像进行中值滤波、图像阈值分割和二值化等一系列预处理,提出了阈值下限的合理选取值,然后对基于Haar-like特征的AdaBoost学习分类器进行使用优化,提出应首先对夜间车辆数据集进行学习分类训练,利用学习后的深度学习机来识别可能作为车辆的假定目标,并把识别为假定车辆的区域标记为感兴趣区域。根据车辆的颜色和几何性质,利用先验知识,对假定为车辆的目标进行确认。最后在Matlab上对整个识别过程进行了验证。
实验结果表明:通过改进阈值下限进行区域分割,确定感兴趣区域,能够提高计算速度;对深度学习分类器数据集进行分类,再进行验证,与以往算法相比,能够提高车辆检测识别率,降低汽车的错误检测率。对于辅助驾驶系统之中的前方车辆识别系统以及汽车防碰撞系统,可以提高汽车的安全性,增强车辆识别系统的准确性。
本文的特色:(1)发现阈值下限对感兴趣区域划分更重要,并提出了优化数值;(2)针对车辆识别部分,提出一种优化学习分类器数据集的算法,实验证明有更高检测率;(3)基于深度学习、机器学习的科学理论,应用了Microsoft Visual Studio 2010,MATLAB R2016a等开发工具,还用了OpenCV机器视觉库并进行实验仿真运行,在实验的基础上,验证了夜间车辆识别系统的可靠性。
关键词:深度学习;车辆识别;Haar-like特征;AdaBoost
Abstract
With the development and progress of the society, our car ownership has been steadily improved. The increase in the number of cars also creates a series of social problems, such as traffic jams, environmental pollution and frequent traffic accidents. Especially at night, the incidence of traffic accidents is much higher than during the day, in order to improve traffic safety coefficient, enhanced auto safety performance, the night in front of the vehicle identification technology research for automatic driving system, the development of automatic identification is of great importance.
This paper is based on the former to the monocular camera image data of the vehicles ahead are extracted, and based on the deep learning technical identification of the vehicles ahead at night. Paper, first of all, the research status of vehicle identification technology at home and abroad, and the development of vehicle identification technology comprehensive elaboration, thought the night light is weaker, vehicle recognition rate are greatly influenced by environment, the speed, the equipment, therefore, the paper selected taillight extraction method to identify the vehicles, first of all, the extraction of vehicle image median filtering, image threshold segmentation and binarization and a series of pretreatment, puts forward the reasonable selection of the lower limit value threshold, and then based on the characteristics of Haar - like AdaBoost learning classifier using optimization, proposed should first to learning classification training of vehicle data set at night, use the depth after learning machine learning to identify possible as assumed targets of the vehicle, and the identification of putative areas marked as interested in areas of the vehicle. According to the color and geometry of the vehicle, use prior knowledge to confirm the target of the vehicle. Finally, the identification process was verified by Matlab.
The experimental results show that the area of interest is determined by improving the threshold of the threshold, and the calculation velocity can be increased. For deep learning classifier to classify data sets, and then validated and compared with the previous methods, can improve recognition rate, vehicle detection reduce error detection rate of the vehicle. For the front of the auxiliary driving system of vehicle identification system, and car collision system, can improve the safety of the car, enhance the accuracy of the vehicle identification system.
The characteristics of this paper are: (1) it is important to find the lower limit to the area of interest, and put forward the optimization value. (2) an algorithm for optimizing the study classifier data set is proposed for vehicle identification. (3) based on the deep learning, machine learning science theory, the application of Microsoft Visual Studio 2010, development tools such as MATLAB R2016a, also using the OpenCV library of machine vision and experimental simulation operation, on the basis of experiment, and verify the reliability of the vehicle identification system at night.
Key words: Depth learning;vehicle identification;Haar-like features;AdaBoost
目 录
第1章 绪论 1
1.1 研究的背景和意义 1
1.2 国内外研究现状 2
1.2.1 车辆识别算法的研究现状 2
1.2.2 车辆识别技术的研究现状 3
1.3 研究的主要内容 4
1.4 论文的组织结构 5
1.5 小结 7
第2章 基于深度学习的车辆感兴趣区域的提取 8
2.1 引言 8
2.1.1 提取感兴趣区域的方法 8
2.1.2 深度学习的介绍 9
2.2 图像的预处理 9
2.2.1 数据集的获取 9
2.2.2 滤波去除噪点 9
2.2.3 图像阈值分割 12
2.3 Haar-like特征 14
2.3.1 Haar-like特征的简介 14
2.3.2 特征值积分图 16
2.4 AdaBoost分类器算法与设计 18
2.4.1 AdaBoost算法 18
2.4.2 AdaBoost算法特征选择 19
2.5 利用学习分类器做假定目标的筛选 21
2.5.1 分类器正训练集的特征提取 22
2.5.2 分类器负训练集的获取 22
2.6 小结 23
第3章 基于先验知识的感兴趣区域车辆确认 24
3.1 坐标系的校准与几何尺寸验证 24
3.2 车大灯的验证 25
3.3 制动灯活动的识别 27
3.4 小结 28
第4章 实验结果的验证分析与评价 29
4.1 本文的实验环境 29
4.2 实验数据集的获取 29
4.2.1 LISA-Night数据集的照明条件 29
4.2.2 数据集 30
4.2.3 数据集的区域标记 31
4.3 使用AdaBoost学习分类器进行训练 31
4.3.1 样本的建立 31
4.3.2 对正样本图片进行归一化处理 35
4.3.3 Adaboost分类器的训练 36
4.3.4 训练结果 37
4.4 性能评估 37
4.5 小结 38
第5章 结论 39
5.1 研究总结 39
5.2 研究展望 39
参考文献 41
致谢 43
第1章 绪论
1.1 研究的背景和意义
伴随着社会的发展进步,经济增长速度不断加快,也带动了汽车产业的发展与繁荣。在已经过去的十几年中,我国的汽车产量以及销量经历了爆发式的增长,汽车产量的增长也推动了我国汽车市场上汽车保有量的显著提升。从公安部交通安全管理局网站获得的数据显示,截至到2014年12月,我国已经保有汽车1.54 亿辆,在世界上居第二位;我们去查阅2015年的国民经济统计公报,会发现全国的汽车保有量也已经达到了惊人的1.72亿辆;根据数据预测,到21世纪20年代以后,我国的民用汽车的总数将会超过2.80亿辆。与此相关的是,随着我国汽车保有量的增长,汽车市场的繁荣,汽车的大众化也给我国的能源、交通和环境造成了许多的问题。但是由于我国的道路交通基础设施的建设速度还没有跟上机动车的增长的速度,因此在社会的建设与发展中引发了环境污染、能源浪费、交通阻塞以及交通事故等一系列的问题。