Effective multimedia surveillance using a human-centric approach

Large-scale multimedia surveillance installations usually consist of a number of spatially distributed video cameras that are installed in a premise and are connected to a central control station, where human operators (e.g., security personnel) remotely monitor the scene images captured by the cameras. In the majority of these systems the ratio of human operators to the number of camera views is very low. This potentially raises the problem that some important events may be missed. Studies have shown that a human operator can effectively monitor only four camera views. Moreover, the visual attention of human operator drops below the acceptable level while performing the task of visual monitoring. Therefore, there is a need for the selection of the four most relevant camera views at a given time instant. This paper proposes a human-centric approach to solve the problem of dynamically selecting and scheduling the four best camera views. In the proposed approach we use a feedback camera to observe the human monitoring the surveillance camera feeds. Using this information, the system computes the operator's attention to the camera views to automatically determine the importance of events being captured by the respective cameras. This real-time non-invasive relevance feedback is then augmented with the automatic detection of events to compute the four best feeds. The experiments show the effectiveness of the proposed approach by improving the identification of important events occurring in the environment.