【作者】 付晓薇；

【导师】 方康玲；

【作者基本信息】 武汉科技大学 ， 计算机应用， 2003， 硕士

【摘要】 动态图像多目标识别与计数，在医学研究、交通监视、客流量统计、天文观测等领域有着非常重要的实用价值和广阔的发展前景。利用图像处理和模式识别技术对运动多目标进行跟踪计数是一种先进的计数手段，是目前国内研究的一个热点。 动态图像多目标识别与跟踪计数是图像处理的一个重要方面。它包括了运动目标的检测、目标分类与提取、目标跟踪与计数几方面，涉及到图像处理与模式识别领域许多核心课题。本论文以“商场客流量图像统计系统”项目为背景，研究了一类运动多目标跟踪计数问题。在本算法中，为了尽可能避免目标的重叠或覆盖的现象，将CCD摄像机安装在获取图像的正上方。并在实验室模拟了一类简单的进出商场人运动情况的动态目标识别与跟踪计数系统。 在本算法完成系统初始化后，首先，利用本文提出的一种改进的自适应运动目标检测方法，检测运动目标是否存在。该方法能有效抑制光照变化、人影、噪声等影响；其次，采用较有效的图像滤波、图像分割、形态学处理等方法对图像进行处理。并根据统计特征清除了伪目标，提取出人头部真正目标，选择了能反映目标不变特征；最后，在比较几种常用多目标跟踪算法的基础上，本算法采用特征跟踪方法对该类多目标进行跟踪计数。跟踪计数是本算法的一个难点。根据该类多目标在相邻帧间运动具有连续性，并且包围窗口变化不大的特点，本跟踪计数算法改进了一种代价函数，应用于图像跟踪的匹配过程中。启动了卡尔曼滤波，预测目标匹配搜索区域。同时，使用目标链记录了目标最新的运动状态和特征值，保证了运动跟踪的连续性。在跟踪过程中考虑了目标暂时静止、消失、交叉多种情况，并对动态图像中帧间出现新目标进行了正确处理，实现了多目标的正确的跟踪，进而对运动目标是否在跟踪区域徘徊进行判断，保证了进出目标计数的正确性。 通过大量实验，本算法对该类动态图像多目标的跟踪计数取得较好效果，计数结果基本上与实际相符。更多还原

【Abstract】 Multi-object recognition and counting in dynamic image has the very important applied value and prospective scene in such fields as the research of the medicine, the transportation monitor statistics of passenger volume and observation of the astronomy. It’s a kind of advanced counting ways to adopt the technology of the image processing and pattern recognition to track and count the moving multi-target. At present, the technology has been become one of research hotspots.Multi-object recognition and counting in dynamic image is an important aspect of the image processing which includes detection classification feature extraction tracking and counting of the moving objects. And many kernel subjects are involved in the fields of image processing and pattern identification.This paper is based on the project "the statistic system of the passenger volume in the shop." A tracking and counting problem is researched on a kind of moving multi-object. The CCD camera is placed above the preyed image so that the overlap phenomenon of the objects can be avoided in the algorithm.Firstly, an advanced approach on moving objects is put forward in this paper, and it is used to detect whether moving objects exist. To a certain degree, this approach can restrain the affection such as the variant shinning shadow noise and so on.Secondly, an effective approach is adopted such as the image filtering, image segmentation, morphology processing and so on. According to the statistic features, false objects are cleared and the true objects of the heads are extracted to choose the invariable features of the objects.Finally, on the base of the normal algorithms about multi-object tracking, the feature-based tracking approach is adopted to track and count the kind of multi-object in the algorithm, in which the tracking and counting is one of the difficulties. Regarding of the moving continuity and the small change of object’s features in frames, a kind of cost function is put forward in the tracking and counting algorithm, which is applied in the tracking match of the image. At the same time, the Kalman filter is used to predicate the search areas of the matching objects, and then,moving state and feature value of the objects in current frame are recorded to ensure the continuity of the dynamic tracking depended on the object-chain. This method can reduce the searching areas of the objects. In the process of tracking, when the objects rest disappear cross, or the new objects appear between the frames in the dynamic image, the algorithm can solve it correctly to realize the proper tracking of the multi-object, what’s more, it correctly judges whether dynamic objects wander in the tracking areas, which ensures the validity of the counting.Based on a large number of experiments, the algorithm has a good effect on tracking and counting of a kind multi-object in dynamic images and counting results coincides with the fact basically.更多还原