【作者】 朱伟；

【导师】 路林吉；

【作者基本信息】 上海交通大学 ， 控制理论与控制工程， 2009， 硕士

【摘要】 随着计算机视觉的发展,人体跟踪技术已经成为一个热点课题,被越来越广泛地应用于生产实际当中。单人体的跟踪技术已经取得了很大的发展,跟踪精度和速度都达到了实用的阶段。然而,多人体跟踪技术却一直是个难点,原因是当多个目标被跟踪时,容易发生目标自遮挡、目标之间的相互遮挡以及背景对目标的遮挡,从而造成跟丢或错跟的情况。国内外学者目前在处理多人体跟踪时,多采用轨迹匹配或者建立运动方程来预测目标的状态,当目标的运动轨迹较为简单时,这些方法可以取得很好的效果。不过,在大多数情况下,目标的运动方向和速度都是不确定的,轨迹和运动状态都难以预测,因此,现有的方法不能很好地解决遮挡问题。本文提出一种新的基于人体匹配的算法来解决多人体跟踪中的遮挡问题。在跟踪目标的同时,分割提取每个人的上半身图像,建立训练样本库。当目标被遮挡时,原先每个目标的人体信息已经记录在样本库中。如果检测到新的目标,则将该目标同原来的样本库进行匹配,如果能匹配上,认为是目标在遮挡后重新出现了,如果不能匹配上,则认为是有新的目标出现,建立新的跟踪轨迹。本文的方法避免了对运动轨迹进行预测,因此,即使被跟踪目标的运动是随机的,也能取得很好的关联效果,使得单人体跟踪可以自由扩展到多人体跟踪上。通过对大量的视频序列进行测试,我们可以发现,相比于其它方法,本文的算法在关联精度上有了很大的提高,较好地解决了多人体跟踪中的遮挡问题。同时,本文将两种特征运用于人体匹配,取得了很好的效果,这对于以后的人体识别技术也有一定的参考价值。更多还原

【Abstract】 With the development of computer vision, human tracking technology has become a heated topic, which has been increasingly widely used in practice. Single-human tracking technology has achieved great development, tracking accuracy and speed has reached a practical stage. However, multi-human tracking technology is still a difficult question, because when multiple targets being tracked, self occlusion, mutual occlusion and background occlusion often happen, resulting in lost tracking or wrong tracking.Presently, scholars at home and aboard deal with multi-human tracking by trajectory matching or dynamic motion equation to predict the target’s state. If the target’s trajectory is relatively simple, these methods can obtain good results. However, in most cases, the movement direction and speed of the target is uncertain, and the trajectory and state are difficult to predict, therefore, the existing methods can not solve the issue of the occlusion.This paper presents a new algorithm based on human matching to solve the occlusion problem in multi-human tracking. At the same time of tracking, each person’s upper body image is extracted, and a training sample pool is created. When the target is occluded, the original body information has been recorded in the sample pool. If a new object is detected, then matching the target with the sample pool. If they match, we consider that the object belongs to the original targets, if not, it is a new object, and a new trajectory should be created.This method avoids the trajectory prediction; therefore, even if the target’s movement is random, the method will also provide a good correlation result. It will let single human tracking extend to multi-human tracking freely.Through a large number of video sequences for testing, we can see that, compared to other methods, the paper’s algorithm provides a better performance on correlation accuracy and a better solution to the occlusion problem in multi-human tracking. Meanwhile, two features have been applied to human matching and achieved good results, which is valuable in human recognition in the future.更多还原