【作者】 陈卓；

【导师】 刘家锋；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2010， 硕士

【摘要】 人体运动跟踪是人体运动分析中一个积极而又重要的研究领域,而人体关节点跟踪又是人体运动跟踪的重要组成部分。本文的研究内容正是跟踪人体全身14个关节点。本文首先建立人体骨架二维模型进行尝试性实验跟踪人体全身关节点,基于该模型的跟踪方法使用粒子滤波器来完成跟踪,而利用粒子滤波在进行多目标跟踪时极易产生“粒子爆炸”问题,由此引入分块采样的策略来优化跟踪算法。而在单纯使用颜色信息进行直方图匹配时,容易产生因为各粒子之间过于自由而出现的“收缩”现象,本研究工作中独创性的提出利用人体生理信息的自定义肢体距离约束,通过该约束来限制并减小这种现象的发生。尽管如此,基于该模型的跟踪头部误差还是很明显的,而且这种误差在自顶向下的跟踪策略中不断累积并传播,无形中增大了其他关节点的跟踪误差。由此激发出一个对头部进行特殊跟踪处理的想法。接下来,本文提出了一个基于三维模型的头部跟踪方法。该跟踪方法首先仍是建立可以反映人体头部生理结构特征不同于任何其它模型的头-肩三维模型,三维模型向平面投影即可获得二维图像模板,基于头-肩三维模型的方法正是通过匹配图像观测与二维模板而完成跟踪的。该跟踪算法无需任何初始化过程,在匹配之前,对图像观测通过背景差分、二值化及边缘检测的方法来获得包括人体头部区域的ROI区域;匹配时利用边缘信息构造相似度计量函数。最后,本文将头-肩三维模型整合到骨架二维模型中构造混合双模型,完成基于混合双模型的人体关节点跟踪。其中,利用头-肩三维模型的跟踪方法跟踪头和双肩关节点,将跟踪结果直接用于骨架二维模型对于全身其他关节点的跟踪,该混合双模型充分吸取了头-肩三维模型基于模板匹配算法的计算速度快和跟踪精度高的优点,同时保留了骨架二维模型对于全身关节点跟踪的应用性。更多还原

【Abstract】 Human motion tracking is an activated an important research domain in human motion analysis. Meanwhile, human joint tracking is an important part of Human motion tracking. The subject of this paper is just tracking all the 14 joints of human body.First, this paper tries to do the tracking job directly by developing a two-dimensional articulated body model, based on which, Particle filtering is used to finish this task. But when Particle filtering is used to do multi-target tracking, it’s to encountering the problem of "particle-explosion". In order to solve the problem, the tactic of partitioned sampling is adopted to optimize the tracking algorithm. If the tracking method only use color histogram matching, the "shrinking" problem which is caused by the freedom of the particles from two different sets arises. In this research, customized physical distance constraints, which can reflect the physical structure information, is proposed originally to limit and even avoid the "shrinking" problem. However, the error of head is so clear that it can accumulate constantly in the top-down tracking tactic. Potentially, the tracking error of the other joints is increased in the process, which encourages an idea of a special head track processing.Then, this paper proposes a head tracking method based on three-dimensional model. The method also need to build a three-dimensional head-shoulder model, which can also reflect the head structure information of human and is different to any other human body. Two-dimensional image template is got by planar projection of the three-dimensional model. And the tracking method is just done by matching the observation and the two-dimensional template. This tracking algorithm doesn’t need any initialization. Before matching, by background substraction, binarization and edge detection on the observation, it’s easy to obtain ROI which comprises the head region of human. When matching, edge information is used to construct likelihood function.At last, this paper builds mixture double model by introducing the three-dimensional head-shoulder model to the two-dimensional articulated model and based on the new model, fulfills the joint tracking of all the body. In which, the head and the two shoulders is tracked by three-dimensional head-shoulder model, and then the tracking result of the head and the two shoulders is used as constant to assist the two-dimensional articulated model for the tracking of other joints. Abundantly, The mixture double model absorbs the advantage of fast computing and high accuracy of the tracking method based on three-dimensional head-shoulder model which uses template matching algorithm. Meanwhile, it inherits the availability of applying the joint tracking of the whole body.更多还原