【作者】 段萌远； 于俊清； 王锦；

【Author】 DUAN Meng-Yuan,YU Jun-Qing, WANG Jin (Department of computer science & technology, Huazhong University of Science & Technology, Wuhan 430074,China)

【机构】 华中科技大学计算机科学与技术学院数字媒体研究室；

【摘要】 针对于固定单摄像机的多人跟踪,提出一种基于活动状态分类的多目标跟踪算法,利用卡尔曼滤波对遮挡进行预测,同时结合区域匹配信息对目标的活动状态进行分类。算法通过采用不同的目标定位及模板更新策略处理不同活动状态的目标,提高了对遮挡跟踪的鲁棒性。实验结果表明,目标的分类处理使算法对多目标跟踪具有较好的适应性和准确性。更多还原

【Abstract】 Moving object tracking, especially tracking of non-rigid objects in digital video surveillance system is a challenging task because a object can rotate, be occulted by other objects or background in image sequence or a object’s size is changed by camera zooming. A multiple objects tracking algorithm is introduced based on object activity classification. It is assumed that there is one static camera in system. The aim is to track multiple people and get their motion trajectory. The whole algorithm is composed of five parts: motion region detection, region matching, object activity classification, object locating, object model update and status prediction. Firstly, connected motion region need to be gotten in the current frame. The background subtraction approach use Gaussian mixture models, which is better to make the background subtraction robust to illumination. The motive foreground regions of each frame are grouped into connected components by analyzing connectedness. And small components can be removed using a size filter. We assumed that occluded objects from the first time have not appeared, so we can get the whole feature information of each object. The color distributing histogram, which includes color spatial information, is used as main feature of objects. It has many advantages for tracking non-rigid objects and is calculated efficiently. For each successive frame, region matching attempts to associate the foreground regions with one of the existing tracked object. This is achieved by constructing a matrix to represent the correlation between each of the motion regions in the current frame and all the tracked objects in the last frame. Kalman prediction equation is used to predicted position information. Here we can get initial position of objects, and new objects and false objects can be distinguished through analyzing region matching result. Predicted position information from Kalman filter is used not only to match region in the next frame but also to predict occlusion status in the next frame between each two objects. Then, using occlusion prediction information, the tracked objects can be classified into several classes. The tracked object can be in different status, such as a new object, moving separately, moving abnormally separately, moving under occlusion, a static object or leaving. Objects in different status are processed in the different locating method. Separate objects can be located exactly only through region matching, and occluded objects have to be located using mean-shift algorithm. In that way, we can exact position coordinates. Finally object models are updated in the proper way according to the similarity between the features of the current objects and former objects. At the same time the Kalman gain is computed and the measurement equation is updated. Experimental results show the good adaptability of the new algorithm to multiple objects tracking. When objects under occlusion are keeping moving, the更多还原