基于光流的动态场景中运动车辆检测与跟踪算法研究
发布时间：2018-11-16 17:08
【摘要】：运动车辆的检测与跟踪技术一直以来都是智能交通系统中的一个重点研究内容,目前在车辆的检测和跟踪问题上还有着很多有待解决的问题。其中动态场景中运动车辆的检测由于存在着车辆与背景两个相互独立的运动而使车辆的提取更加困难,在车辆检测的准确性上,有着很大的研究空间。论文首先对目前的光流检测方法中存在的错误光流难以完全去除的问题进行了详细的研究和探讨,并提出了一种有效的解决方法。同时采用聚类的方法准确的从动态场景中提取出车辆。最后研究了车辆跟踪技术,对车辆的遮挡问题进行了详细的讨论,并通过算法的改进,有效的解决了此问题。 在车辆检测的问题中,采用Harris算子计算图像的特征点,再通过金字塔Lucas-Kanade光流(L-K光流)法计算图像的特征点光流场。随后引入矢量量化的思想对图像的光流场进行聚类,并提出了结合欧式距离和相似系数作为相似性测度的方法提高了聚类的准确性。最后计算各个类别中角点的分布方差,通过RANSAC方法对光流场中的错误光流进行粗剔除,再依据类内方差值进行精剔除。最后再依据类内方差值的大小实现车辆的提取。 在车辆跟踪的问题中,深入研究了Camshift和Kalman滤波相结合的方法在车辆跟踪中的应用,并对目前的跟踪算法中普遍存在的遮挡问题进行了详细的讨论,最后通过增加区域面积约束,对算法进行改进,提高了车辆在被遮挡和有颜色干扰的情况下的跟踪的准确性。 通过具体的实验表明,检测算法能够准确的从动态场景中检测出运动的车辆,对单个运动车辆的检测准确度可达93%,多车辆的情况下,在车辆数量较少且较为分散的情况下能达到80%,具有较高的准确性。同时,算法也表现出了很好的跟踪效果,在车辆被遮挡时间较短的情况下,仍然实现能够稳定的跟踪。且算法基本上能达到实时性的要求。
[Abstract]:The detection and tracking technology of moving vehicles has always been an important research content in Intelligent Transportation system. At present, there are still many problems to be solved in the detection and tracking of vehicles. The detection of moving vehicles in dynamic scene makes it more difficult to extract vehicles due to the existence of two independent movements of vehicle and background. There is a great space for research on the accuracy of vehicle detection. In this paper, the problem that the wrong optical flow is difficult to be completely removed in the current optical flow detection methods is studied and discussed in detail, and an effective solution is proposed. At the same time, the method of clustering is used to extract the vehicle from the dynamic scene accurately. Finally, the vehicle tracking technology is studied, and the occlusion problem is discussed in detail, and the algorithm is improved to solve the problem effectively. In the problem of vehicle detection, the Harris operator is used to calculate the feature points of the image, and then the image characteristic point optical flow field is calculated by the pyramid Lucas-Kanade optical flow (L-K optical flow) method. Then the idea of vector quantization is introduced to cluster the optical flow field of the image, and the accuracy of clustering is improved by combining Euclidean distance and similarity coefficient as similarity measure. Finally, the distribution variance of corner points in each category is calculated, and the error optical flow in the optical flow field is eliminated rough by RANSAC method, and then refined elimination is carried out according to the intra-class variance value. Finally, the extraction of vehicles is realized according to the value of intra-class variance. In the problem of vehicle tracking, the application of Camshift and Kalman filtering in vehicle tracking is deeply studied, and the common occlusion problem in the current tracking algorithm is discussed in detail. Finally, the area constraint is added to the tracking algorithm. The algorithm is improved to improve the accuracy of vehicle tracking under occlusion and color interference. The experimental results show that the detection algorithm can accurately detect moving vehicles from the dynamic scene, and the detection accuracy of single moving vehicle can reach 933%, in the case of multiple vehicles, When the number of vehicles is small and scattered, it can reach 80% and has high accuracy. At the same time, the algorithm also shows a good tracking effect, in the case of vehicle occlusion time is short, still can achieve stable tracking. And the algorithm can basically achieve the requirement of real-time.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495
本文编号：2336110
[Abstract]:The detection and tracking technology of moving vehicles has always been an important research content in Intelligent Transportation system. At present, there are still many problems to be solved in the detection and tracking of vehicles. The detection of moving vehicles in dynamic scene makes it more difficult to extract vehicles due to the existence of two independent movements of vehicle and background. There is a great space for research on the accuracy of vehicle detection. In this paper, the problem that the wrong optical flow is difficult to be completely removed in the current optical flow detection methods is studied and discussed in detail, and an effective solution is proposed. At the same time, the method of clustering is used to extract the vehicle from the dynamic scene accurately. Finally, the vehicle tracking technology is studied, and the occlusion problem is discussed in detail, and the algorithm is improved to solve the problem effectively. In the problem of vehicle detection, the Harris operator is used to calculate the feature points of the image, and then the image characteristic point optical flow field is calculated by the pyramid Lucas-Kanade optical flow (L-K optical flow) method. Then the idea of vector quantization is introduced to cluster the optical flow field of the image, and the accuracy of clustering is improved by combining Euclidean distance and similarity coefficient as similarity measure. Finally, the distribution variance of corner points in each category is calculated, and the error optical flow in the optical flow field is eliminated rough by RANSAC method, and then refined elimination is carried out according to the intra-class variance value. Finally, the extraction of vehicles is realized according to the value of intra-class variance. In the problem of vehicle tracking, the application of Camshift and Kalman filtering in vehicle tracking is deeply studied, and the common occlusion problem in the current tracking algorithm is discussed in detail. Finally, the area constraint is added to the tracking algorithm. The algorithm is improved to improve the accuracy of vehicle tracking under occlusion and color interference. The experimental results show that the detection algorithm can accurately detect moving vehicles from the dynamic scene, and the detection accuracy of single moving vehicle can reach 933%, in the case of multiple vehicles, When the number of vehicles is small and scattered, it can reach 80% and has high accuracy. At the same time, the algorithm also shows a good tracking effect, in the case of vehicle occlusion time is short, still can achieve stable tracking. And the algorithm can basically achieve the requirement of real-time.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495
【参考文献】
相关期刊论文 前10条
1 史新宏,蔡伯根,穆建成;智能交通系统的发展[J];北方交通大学学报;2002年01期
2 施家栋;王建中;王红茹;;基于光流的人体运动实时检测方法[J];北京理工大学学报;2008年09期
3 宋寅卯;李晓娟;刘磊;;图像噪声滤波的研究方法及进展[J];电脑开发与应用;2010年04期
4 李俊韬,张海,范跃祖,王力;复杂场景条件下的运动目标检测算法[J];光电工程;2004年S1期
5 李智华,王玉文;卡尔曼滤波在图象识别中的应用[J];哈尔滨师范大学自然科学学报;2005年03期
6 王宏,何克忠,张钹;智能车辆的自主驾驶与辅助导航[J];机器人;1997年02期
7 杨建荣;曲仕茹;;基于单目视觉的障碍物检测方法研究[J];计算机仿真;2009年02期
8 郭庆昌;蔡劏;郭盛雨;;基于均值移动算法的图像分割的研究[J];计算机仿真;2010年02期
9 郭磊;李克强;王建强;连小珉;;一种基于特征的车辆检测方法[J];汽车工程;2006年11期
10 周志宇,汪亚明,黄文清,曹丽,周海英;基于熵的运动目标检测[J];微计算机信息;2003年05期
，本文编号：2336110
本文链接：https://www.wllwen.com/kejilunwen/jiaotonggongchenglunwen/2336110.html
