基于外观特征和交互信息的多目标跟踪
发布时间：2019-03-12 16:02
【摘要】：现代社会,无处不在的摄像设备等使得视频数据急剧增加,依靠人工处理的方式已无法满足实际需求。如何采用机器处理的方式,高效、准确地从这些视频数据中获取需要的信息已成为计算机视觉领域重要的研究任务。而目标跟踪,尤其是多目标跟踪已成为这些任务中重要且基础的一个研究方向。在公共安全、智能交通等视频数据的处理时,都需要先获取监测到的各个目标的运行轨迹,再判断是否有公共事件的发生或判断交通违章情况。此外,在人机交互,自动驾驶等场景下,目标跟踪也有广泛的应用。在对多个目标进行跟踪的时候,需要面对各种复杂的情况,比如需要在目标被遮挡、轨迹交叉、光照变化等情况下将不同的目标分辨开来,持续跟踪下去。现在最常用的多目标跟踪的方法是先检测后跟踪的方法。先采用检测算法对图像区域进行检测获取检测结果,再利用跟踪算法对目标进行持续的跟踪,将检测结果关联成不同目标的轨迹。在对目标进行跟踪的时候,需要利用目标的视觉信息,即目标的各种特征。对这些特征的要求首先是具有良好的分辨能力,能够很好的将属于不同目标的检测结果分辨开;同时还要求其具有连续性,即在不同帧的属于同一个目标的检测结果的特征要尽可能的相似。而现有的单一特征或单一种类特征的跟踪方法很难同时满足上述两点要求。在目标跟踪过程中也经常出现跟踪的目标被遮挡等情况,使目标的视觉特征无法获取,或目标外观发生变化等使单纯依赖视觉特征的跟踪无法继续。而在跟踪目标为人的时候,例如行人跟踪,多个目标之间的运动模式是相互影响的,周围目标的运动模式有助于对当前目标的跟踪,即目标的交互信息可以提高跟踪的性能。针对以上分析,本文的主要研究内容和创新点主要有以下两点:1.提出了一种基于全局和局部特征的实时多目标跟踪的方法。该方法是一种分两步的,分别利用了全局特征和局部特征的同时对多个目标进行跟踪的方法。全局跟踪阶段采用全局特征,即广泛应用的颜色直方图以满足对目标特征连续性的要求;局部跟踪阶段采用了一种改进的基于最大稳定极值区域的局部特征,满足对目标特征分辨能力的要求。这样便同时满足跟踪中对分辨能力和连续性的要求,且拥有比较低的计算复杂度,可以应用于实时跟踪。2.提出了一种基于队列稳定性的多目标跟踪方法。在F-Formation的基础上提出了队列稳定性的概念并将队列稳定性用于提高半拥挤环境下多目标跟踪的性能,且将其嵌入传统跟踪框架中。在跟踪的过程中,在利用传统的视觉信息等之外,还需要保持两段轨迹片段在关联前后整个队列的稳定性,使其不发生大的变化,提高了跟踪的性能。
[Abstract]:The modern society, the ubiquitous imaging equipment and so on make the video data increase dramatically, and the actual demand cannot be met by means of manual processing. How to adopt a machine-based approach to efficiently and accurately obtain the required information from these video data has become an important research task in the field of computer vision. Target tracking, especially multi-target tracking, has become an important and fundamental research direction in these tasks. In the process of processing video data such as public safety and intelligent traffic, it is necessary to first obtain the running track of each target monitored, and then judge whether there is a public event or judge the traffic violation condition. In addition, under the scene of man-machine interaction and automatic driving, the target tracking has a wide application. When tracking a plurality of targets, it is necessary to face a variety of complex situations, such as the need to distinguish different targets in the case of a target being blocked, a track crossing, a light change, and the like, and continue to follow. The most commonly used method of multi-target tracking is to detect post-tracking methods first. Firstly, a detection algorithm is adopted to detect the image area to obtain a detection result, and the target is continuously tracked by a tracking algorithm, and the detection result is correlated to a track of different targets. When tracking a target, it is necessary to take advantage of the visual information of the target, that is, the various features of the object. The requirements for these features are first to have good resolving power and to distinguish the detection results belonging to different targets well, and to have continuity, that is, the characteristics of the detection results belonging to the same object in different frames are to be as similar as possible. And the existing single-feature or single-type feature tracking method is difficult to meet the two-point requirements simultaneously. In the target tracking process, the target of the tracking is blocked and the like, so that the visual characteristics of the target can not be acquired, or the appearance of the target changes and the like, so that the tracking of the pure-dependent visual features can not be continued. While tracking the target as a person, such as pedestrian tracking, the motion pattern between the plurality of targets is interactive, and the motion pattern of the surrounding target contributes to the tracking of the current target, that is, the interaction information of the target can improve the performance of the tracking. For the above analysis, the main research content and innovation point of this paper have the following two points:1. A method for real-time multi-target tracking based on global and local features is presented. The method is a two-step method for tracking a plurality of targets at the same time using global features and local features, respectively. The global tracking stage adopts the global feature, that is, the widely used color histogram is used to meet the requirement of the continuity of the target feature; the local tracking stage adopts an improved local feature based on the maximum stable extreme value region, and meets the requirement of the target feature resolution capability. In that way, the requirements for resolution and continuity in the tracking are met at the same time, and the computational complexity of the tracking is relatively low, and the method can be applied to real-time tracking. A multi-target tracking method based on queue stability is proposed. On the basis of F-Formation, the concept of queue stability is proposed and the stability of the queue is used to improve the performance of multi-target tracking in a semi-crowded environment, and it is embedded in the traditional tracking framework. In the process of tracking, in addition to the traditional visual information and the like, the stability of the whole queue before and after the association is required to be maintained, so that the stability of the whole queue before and after the association is not changed, and the tracking performance is improved.
【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】：2017
【分类号】：TP391.41
本文编号：2438929
[Abstract]:The modern society, the ubiquitous imaging equipment and so on make the video data increase dramatically, and the actual demand cannot be met by means of manual processing. How to adopt a machine-based approach to efficiently and accurately obtain the required information from these video data has become an important research task in the field of computer vision. Target tracking, especially multi-target tracking, has become an important and fundamental research direction in these tasks. In the process of processing video data such as public safety and intelligent traffic, it is necessary to first obtain the running track of each target monitored, and then judge whether there is a public event or judge the traffic violation condition. In addition, under the scene of man-machine interaction and automatic driving, the target tracking has a wide application. When tracking a plurality of targets, it is necessary to face a variety of complex situations, such as the need to distinguish different targets in the case of a target being blocked, a track crossing, a light change, and the like, and continue to follow. The most commonly used method of multi-target tracking is to detect post-tracking methods first. Firstly, a detection algorithm is adopted to detect the image area to obtain a detection result, and the target is continuously tracked by a tracking algorithm, and the detection result is correlated to a track of different targets. When tracking a target, it is necessary to take advantage of the visual information of the target, that is, the various features of the object. The requirements for these features are first to have good resolving power and to distinguish the detection results belonging to different targets well, and to have continuity, that is, the characteristics of the detection results belonging to the same object in different frames are to be as similar as possible. And the existing single-feature or single-type feature tracking method is difficult to meet the two-point requirements simultaneously. In the target tracking process, the target of the tracking is blocked and the like, so that the visual characteristics of the target can not be acquired, or the appearance of the target changes and the like, so that the tracking of the pure-dependent visual features can not be continued. While tracking the target as a person, such as pedestrian tracking, the motion pattern between the plurality of targets is interactive, and the motion pattern of the surrounding target contributes to the tracking of the current target, that is, the interaction information of the target can improve the performance of the tracking. For the above analysis, the main research content and innovation point of this paper have the following two points:1. A method for real-time multi-target tracking based on global and local features is presented. The method is a two-step method for tracking a plurality of targets at the same time using global features and local features, respectively. The global tracking stage adopts the global feature, that is, the widely used color histogram is used to meet the requirement of the continuity of the target feature; the local tracking stage adopts an improved local feature based on the maximum stable extreme value region, and meets the requirement of the target feature resolution capability. In that way, the requirements for resolution and continuity in the tracking are met at the same time, and the computational complexity of the tracking is relatively low, and the method can be applied to real-time tracking. A multi-target tracking method based on queue stability is proposed. On the basis of F-Formation, the concept of queue stability is proposed and the stability of the queue is used to improve the performance of multi-target tracking in a semi-crowded environment, and it is embedded in the traditional tracking framework. In the process of tracking, in addition to the traditional visual information and the like, the stability of the whole queue before and after the association is required to be maintained, so that the stability of the whole queue before and after the association is not changed, and the tracking performance is improved.
【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】：2017
【分类号】：TP391.41
【参考文献】
相关硕士学位论文 前2条
1 刘文;基于块的多特征目标跟踪算法[D];大连理工大学;2010年
2 柳涛;多通道图像MSER局部不变特征提取算法研究[D];国防科学技术大学;2010年
，本文编号：2438929
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2438929.html
