基于机器视觉的智能驾驶车辆的目标识别研究

发布时间：2018-06-08 23:32

本文选题：智能车辆 + 机器视觉　；参考：《吉林大学》2017年硕士论文

【摘要】：交通一直是人类生活中不可或缺的一部分,伴随着人类生活的方方面面。随着人口总数增加,出行量不断变大、交通压力日益增加、交通事故发生数量也逐年上升。伴随着人工智能科技的不断突破,人们渴望智能交通的出现会给行人出行带来更安心的保障,因此智能交通也给智能汽车的研发和商业化带来了强劲的动力,成为不可逆转的大潮趋势。人们因此有望在未来真正的解放紧握方向盘的双手,来享受智能驾驶带来的安全体验。本文结合某预研项目,定位于面向智能汽车智能辅助驾驶系统环境感知中的目标识别领域,针对当前智能车辆辅助驾驶系统环境感知中目标识别存在的漏检与误检问题,对基于支持向量机(以下简称SVM)分类器的车辆目标和行人目标等研究内容进行了识别研究,并对基于单目视觉的前方运动目标距离进行了理论及试验研究,具体研究内容和结论如下:首先,阐述了智能车辆的研究现状、发展趋势、研究成果、政策方针等相关内容,接着目标识别中机器视觉应用的必要性和广泛性进行探讨,概述了研究背景及意义,并确定了本文的关键技术和研究内容,进而据此进行了章节撰写安排。其次,对图像目标识别技术进行了相关理论研究。对比分析了两种典型机器视觉传感器,确定以CCD摄像机作为本文的研究对象;并对图像灰度化、图像灰度增强、图像滤波、图像形态学等图像预处理技术进行了研究;同时对基于机器学习的目标识别技术进行了研究,确定本文研究中采用适用于大量样本数据集情况下的支持向量机算法。接着,对车辆及行人的特征提取识别技术进行了相关理论研究。概述了特征的概念及其在目标识别过程中所起到的功用;阐释了边缘特征、Hough特征、Surf特征、颜色特征、纹理特征等典型图像特征,重点对HOG特征进行了研究和分析;然后对SVM支持向量机的原理及基本算法进行了研究,确定了基于样本HOG特征信息的车辆目标识别方法及基于样本DPM特征信息的行人目标识别方法。然后,对基于机器视觉的前方车辆距离测量进行了相关研究。比较分析了常用测距传感器的特点,提出了采用基于机器学习的单目视觉距离测量方案;同时对摄像机进行了标定研究,在张氏标定原理基础上,提出了自标定算法的设想;进行基于小孔成像原理的逆过程,根据图像质心像素坐标和摄像机标定内外参数,推导出单目视觉距离测量方程,构建了基于机器视觉的前方车辆距离测量模型。最后,结合前期相关理论研究对研究课题进行了试验验证研究。一方面对车辆目标识别进行了试验验证研究,通过与实景道路环境目标试验对比,验证了基于多核核函数优化的车辆目标识别SVM分类器的有效性和准确性;另一方面对行人目标识别进行了试验验证研究,通过与实景道路环境目标试验对比,验证了基于改进HOG特征的行人目标识别可变形部件模型特征(以下简称DMP)分类器的有效性和准确性;此外,还对基于机器视觉的前方车辆距离测量进行了试验测试,验证了单目视觉距离测量模型的有效性和准确性。本文对智能车辆中环境感知关键技术的研究进行了有益的尝试,其研究成果对于提高识别道路运动目标,改善漏检和误检的识别现象具有一定的现实意义,对于智能车辆环境感知技术中机器视觉技术的应用与推广也有一定的理论和参考价值。
[Abstract]:Traffic has always been an integral part of human life, with all aspects of human life. With the increase of the population, the volume of travel is increasing, traffic pressure is increasing, and the number of traffic accidents is increasing year by year. With the continuous breakthrough of artificial intelligence technology, people are eager for the appearance of intelligent traffic to travel to pedestrians. It brings more security, so intelligent transportation has also brought strong impetus to the development and commercialization of smart cars. It has become an irreversible trend. It is hopeful that in the future, the real liberation of the hands of the steering wheel will be held to enjoy the safety test brought by intelligent driving. In the field of target recognition in intelligent vehicle driving system environment perception, the research content of vehicle target and pedestrian target based on support vector machine (hereinafter referred to as SVM) classifier is identified and researched in view of the problem of missing and mischecking of target recognition in the environment perception of intelligent vehicle auxiliary driving system. The theoretical and Experimental Research on the distance of moving target in front of monocular vision is carried out. The contents and conclusions are as follows: firstly, the research status, development trend, research results, policy and policy of intelligent vehicles are expounded, and then the necessity and universality of the application of machine vision in target recognition are discussed, and the research background is summarized. The key technology and research content of this paper are determined, and then the chapters are arranged. Secondly, the theory of image target recognition is studied. Two typical machine vision sensors are compared and analyzed. The CCD camera is selected as the research object of this paper, and the image grayscale and image gray enhancement are enhanced. The technology of image preprocessing, such as image filtering and image morphology, is studied. At the same time, the technology of target recognition based on machine learning is studied. The support vector machine algorithm is used in this study, which is suitable for a large number of sample data sets. Then, the related theory of feature extraction and recognition for vehicles and pedestrians is carried out. The concept of feature and its function in the process of target recognition are summarized. The typical features such as edge features, Hough features, Surf features, color features and texture features are explained, and the features of HOG are studied and analyzed. Then, the principle and basic algorithms of SVM support vector machines are studied, and the basis is determined. The vehicle target recognition method with sample HOG feature information and the method of pedestrian target recognition based on the sample DPM feature information. Then, the relative research on the distance measurement of the vehicle based on machine vision is carried out. The characteristics of the common range sensor are compared and analyzed, and a single eye vision distance measurement scheme based on machine learning is proposed. The camera is calibrated and the self calibration algorithm is proposed on the basis of the principle of Zhang's calibration. The inverse process based on the principle of small hole imaging is carried out. According to the pixel coordinates of the image center and the camera calibration and the inside and outside parameters, the monocular vision distance measurement equation is derived, and the distance measurement in front of the vehicle based on the machine vision is constructed. On the one hand, the test verification of vehicle target recognition is carried out, and the validity and accuracy of the SVM classifier based on multi kernel kernel function optimization is verified by comparison with the real road environment target test; the other party is proved to be effective and accurate. In the face of pedestrian target recognition, experimental verification is carried out. By comparing with the actual road environment target test, the validity and accuracy of the deformable component model feature (hereinafter referred to as DMP) classifier based on improved HOG features is verified. In addition, the test of vehicle distance based on machine vision is also tested. The test results verify the validity and accuracy of the monocular vision distance measurement model. This paper makes a useful attempt to study the key technologies of environmental perception in intelligent vehicles. The research results have some practical significance for improving the recognition of road moving targets and improving the identification of leakage and misdiagnosis. The application and promotion of machine vision technology in knowledge technology also has certain theoretical and reference value.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U463.6;U495;TP391.41

【相似文献】