Vision Based Calibration and Localization Technique for Video Sensor Networks

The recent evolutions in embedded systems have now made the video sensor networks a reality. A video sensor network consists of a large number of low cost camera-sensors that are deployed in random manner. It pervades both the civilian and military fields with huge number of applications in various areas like health-care, environmental monitoring, surveillance and tracking. As most of the applications demand the knowledge of the sensor-locations and the network topology before proceeding with their tasks, especially those based on detecting events and reporting, the problem of localization and calibration assumes a significance far greater than most others in video sensor network. The literature is replete with many localization and calibration algorithms that basically rely on some a-priori chosen nodes, called seeds, with known coordinates to help determine the network topology. Some of these algorithms require additional hardware, like arrays of antenna, while others require having to regularly reacquire synchronization among the seeds so as to calculate the time difference of the received signals. Very few of these localization algorithms use vision based technique. In this work, a vision based technique is proposed for localizing and configuring the camera nodes in video wireless sensor networks. The camera network is assumed randomly deployed. One a-priori selected node chooses to act as the core of the network and starts to locate some other two reference nodes. These three nodes, in turn, participate in locating the entire network using tri-lateration method with some appropriate vision characteristics. In this work, the vision characteristics that are used the relationship between the height of the image in the image plane and the real distance between the sensor node and the camera. Many experiments have been simulated to demonstrate the feasibility of the proposed technique. Apart from this work, experiments are also carried out to locate any other new object in the video sensor network. The experimental results showcase the accuracy of building up one-plane network topology in relative coordinate system and also the robustness of the technique against the accumulated error in configuring the whole network