动态背景下的行人检测与跟踪研究

发布时间：2019-06-28 18:34

【摘要】：运动目标检测与跟踪是计算机视觉中的一个热点问题,并在智能监控系统中越来越受到人们的重视。根据摄像机与监控背景之间的相对运动关系,将目标运动分为静态背景下的与动态背景下的两类运动。目前,静态背景下的运动目标研究已经相对成熟,并广泛的应用于社会的各个领域,而动态背景下的目标运动相对复杂,理论与实际应用方面还有很多问题需要解决,本文主要对动态背景下的行人目标进行检测与跟踪研究,通过实验验证本文算法的可行性。本文首先对行人检测与跟踪的现状进行了综述,分析了行人检测与跟踪过程中可能遇到的各个方面的问题,并结合数字图像处理的知识,分析了目标检测在静态背景与动态背景两种情况下的主要提取方法。动态背景下目标的运动会受到相机运动产生的全局运动的干扰,需要我们对相机的运动进行背景补偿。本文通过SIFT特征点进行点匹配,完成仿射参数模型的参数估计,在估计过程中,提出了一种新的间隔匹配的方法去除特征点中外点的干扰。在目标提取中,采用多帧差分的方法增强目标检测的判别能力,很好的满足了检测的需要。针对行人目标的跟踪问题,本文首先介绍了几种传统的目标跟踪方法,分析了几种算法的不足之处,并应用Camshift算法进行验证,得出单一的特征很难在复杂的环境下实现目标的描述,容易发生目标丢失的情况。本文提出了多特征融合的方法对跟踪算计进行改进,针对不同环境下目标跟踪的需求,可以选择不同的特征进行融合,改进后的跟踪算法能够很好的实现目标的有效跟踪,提高了算法的鲁棒性。动态背景下的目标跟踪具有广阔的应用前景,在智能汽车、无人机拍摄、特种作战等领域都能够得到很好的应用。本文主要对移动设备上实现行人目标的自动检测与跟踪进行了分析,通过实验结果验证系统的可行性,结果表明,本文提出的算法能够满足移动设备上行人自动检测与跟踪的要求,具有一定的实用价值。
[Abstract]:Moving target detection and tracking is a hot issue in computer vision, and more attention has been paid to it in intelligent monitoring system. According to the relative motion relationship between camera and monitoring background, the target motion is divided into two kinds of motion in static background and dynamic background. At present, the research of moving target in static background has been relatively mature and widely used in various fields of society, while the moving of target in dynamic background is relatively complex, and there are still many problems to be solved in theory and practical application. In this paper, the detection and tracking of pedestrian targets in dynamic background are mainly studied, and the feasibility of the algorithm is verified by experiments. In this paper, the present situation of pedestrian detection and tracking is reviewed, and the problems that may be encountered in the process of pedestrian detection and tracking are analyzed. Combined with the knowledge of digital image processing, the main extraction methods of target detection in static background and dynamic background are analyzed. Under the dynamic background, the motion of the target is interfered by the global motion caused by the motion of the camera, so we need to compensate the background of the motion of the camera. In this paper, the parameter estimation of affine parameter model is completed by point matching of SIFT feature points. in the process of estimation, a new interval matching method is proposed to remove the interference of external points in feature points. In target extraction, multi-frame difference method is used to enhance the discriminant ability of target detection, which meets the needs of detection. Aiming at the problem of pedestrian target tracking, this paper first introduces several traditional target tracking methods, analyzes the shortcomings of several algorithms, and verifies them with Camshift algorithm. It is concluded that a single feature is difficult to achieve the target description in a complex environment, and the target loss is easy to occur. In this paper, a multi-feature fusion method is proposed to improve the tracking algorithm. According to the requirements of target tracking in different environments, different features can be selected for fusion. The improved tracking algorithm can achieve effective target tracking and improve the robustness of the algorithm. Target tracking in dynamic background has a broad application prospect, and can be well applied in intelligent vehicles, UAV photography, special operations and other fields. In this paper, the automatic detection and tracking of pedestrian targets on mobile devices is analyzed, and the feasibility of the system is verified by experimental results. The results show that the algorithm proposed in this paper can meet the requirements of automatic detection and tracking of pedestrians on mobile devices, and has certain practical value.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】