Chinese Journal of Computers   Full Text
Title	Region-SIFT Descriptor Based Correspondence Between Multiple Cameras
Authors	MING An-Long MA Hua-Dong
Address	(Beijing Key Laboratory of Intelligent Telecommunications Software and Multimedia, Beijing University of Posts and Telecommunications, Beijing 100876)
Year	2008
Issue	No.4(650—661)
Abstract &   Background	Abstract This paper proposes a region-SIFT descriptor based target matching method for multiple cameras. This is a region based method. However, the region is represented by SIFT descriptor instead of traditional color features. In the new method, the background subscription is used in multi-target detecting, and the particle filter is utilized in multi-target tracking. Non camera calibrations are required in the new method, neither the constraint that all objects stand in the same plane. The main features of the method are highlighted as follows: (1)Non geometry constraints are required to some extent. (2)Many types of objects are supported. (3)The cameras can simply move during object tracking. (4)It is more fit to distributed computing, but the traditional centralized computing is also supported. (5)It is robust to changes of the light intensity. Experimental results show that the method is effective. keywords region-SIFT descriptor; multiple cameras; correspondence; multi-target detection and tracking Background This work is supported by the National High Technology Research and Development Program of China under grant No.2006AA01Z304; the National Natural Science Foundation of China under Grant No.90612013; the Specialized Research Fund for the Doctoral Program of Higher Education under grant No.20050013010; the NCET of MOE, China. Visual surveillance using multiple cameras has attracted much attention in the computer vision community in resent years. This is because by utilizing multiple cameras, the area of surveillance is expanded and information from multiple views is extremely helpful to handle many issues such as occlusion. However, visual surveillance using multiple cameras also brings a number of problems such as camera installation, calibration of multiple cameras, correspondence between multiple cameras, automated camera switching, and data fusion. Correspondence between multiple cameras involves finding correspondences at the same time between objects in the different image sequences. Only after correspondence between multiple cameras is well constructed can the information from multiple cameras be fused. Therefore, it is one of the most important and basic problems in visual surveillance using multiple cameras. Although correspondence of multiple cameras is a newly emergent research topic, some attempts have been made to investigate this problem. The existing methods for establishing correspondences can be classified according to the types of employed features, whether the cameras are calibrated or not, and whether the correspondences are region-based or point-based. In this paper, we propose a region-SIFT descriptor based target matching method for multiple cameras. This is a region based method. However, the region is represented by SIFT descriptor instead of traditional color features. In our method, the background subscription is used in multi-target detecting, and the particle filter is utilized in multi-target tracking. Non camera calibrations are required in our method, neither the constraint that all objects stand in the same plane. The main features of our method are highlighted as follows: (1) Non geometry constraints are required to some extent. (2) Many types of objects are supported. (3) The cameras can simply move during object tracking. (4) It is more fit to distributed computing, but the traditional centralized computing is also supported. (5) It is robust to changes of the light intensity. Experimental results show that our method is effective.