The Management of a Multicamera Tracking System for Videosurveillance by Using an Agent Based Approach

Abstract

The development of vision systems for monitoring or surveillance of wide area sites is an interesting field of investigation. In order to maximize the capabilities and performance of such system, it is often necessary to use a variety of sensor devices that complement each other. The standard configuration consists in completely covering a scene with a set of cameras with adjacent Fields Of View (FOV). Many people from the computer vision community have worked on the geometrical aspect of this configuration. By using several overlapping calibrated cameras, the system generates a global virtual view of the scene. The use of multiple views of the same scene in the tracking process provides the ability to resolve a part of occlusion situations. A second and less explored configuration is based on a network of non-overlapping cameras. This second configuration is economically attractive because it permits to efficiently decrease the number of sensors. However, the incomplete coverage makes the tracking problem more difficult. The main difficulty is the establishment of correspondence between the objects captured by multiple sensors (cross-camera data association). In this work we present a high level sensor management strategy in a context of videosurveillance including both of the two configurations: overlapping and distant cameras. The global objective of the system is the development of the object tracking task. The general problems of multi-sensor management are related to decisions about what sensors to use and for which purposes, as well as when, where and how to use them. This last side of high level management is closely linked with the concept of active perception strategy (Bajcsy, 1998). This strategy is particularly adapted where real-time performance is needed such as tracking, robot navigation, surveillance, visual inspection. The active perception has been widely developed for designing the perception for mobile robotic. In fact real-time perception systems have theirs limitation in the computation of massive amount of input data with processing procedures in a reduced and fixed amount of time. The active strategy has the capacity to filter data and to focus the attention of the perception to relevant information and also can choose the best alternative by using the contextual information. Such approaches are closely linked with the design of cognitive system which permits to combine knowledge and reasoning in order to develop smart and robust perception system.