Chinese Journal of Computers   Full Text
Title	Multiple Objects Tracking Based on Online Sampling
Authors	ZHU Lin ZHOU Jie SONG Jing-Yan
Address	(Department of Automation, Tsinghua University, Beijing 100084)
Year	2008
Issue	No.1(151—160)
Abstract &   Background	Abstract Multiple objects tracking, especially multiple objects tracking with occlusion, is an important and challenging problem in computer vision. This paper proposes a novel on-line sample based framework to deal with this problem. First, training samples are obtained from each isolated object before occlusion occurs. A simple sample based scheme is used to produce a set of classifiers. The objects in the following frames are labeled using the classifiers. If no occlusion occurs in the frame, the classifiers will be updated by adding new training samples. When there is occlusion in the present frame, according to the labeled results, the foreground region can be segmented into multiple objects. Compared with previous works, this approach needs neither the depth information of the scene nor the prior models of objects such as color blobs, the type of objects (person or vehicle), velocity assumption. Experiments show that the proposed approach can work robustly under the more complex conditions in which the above assumptions may be unreliable. keywords occlusion; multiple objects tracking; online sampling background In video surveillance systems, accurate and real-time track for multiple objects is greatly helpful to object recognition, activity analysis and high-level event understanding. Segmentation and tracking multiple objects is important not only for visual surveillance, but also for other video analysis application such as video indexing, video archival and retrieval system. Many tracking systems now can track the isolated objects well, but may lose the objects for a period of time or even miss the objects through occlusion. So, multiple objects tracking, especially multiple objects tracking with occlusion, is an important and challenging problem. Recently, many researches aim to handle the problem of tracking multiple objects though occlusion by using some prior assumptions, such as, the deep information of the scene known, the kind of objects known(to build the 2D or 3D model), motion regularity. But in real complex situations, the prior assumptions may be unreasonable, which may cause tracking failure. However human being can stably track objects with occlusion by online learning the information of the objects, even if the objects′ kind, motion regularity and the scene depth information are unknown. Inspired by the online learning approach of human vision, the authors propose a novel online sample based framework to track multiple objects with occlusion. First, training samples are online obtained from each isolated object before occlusion to produce a set of classifiers. Then the foreground region through occlusion can be segmented into multiple objects by the classifiers′ labelling. The authors test the proposed framework in many indoor/outdoor data, and the results show that the new approach can track multiple objects with occlusion fast and stably even when the objects′ kind, motion uniform degree and the scene depth information are unknown.