"Video Scene Assesment with an Unattended Sensor Network"
S. Guler, J.A. Silverstein, and K. Garg
SPIE Europe, Security and Defence, Florence, Italy September, 2007 SPIE Europe, Security and Defence, Florence, Italy September, 2007
This paper describes a prototype video analytics based system to monitor remote, and possibly unfriendly, areas with the purpose of being aware and alert of unusual activity. This application involves analysis and metadata extraction from collected multiple video streams, assessment and prioritization of the scene, comprehension of the information and representation of the extracted knowledge. We present a highly efficient video content extraction and scene assessment paradigm with algorithms designed after the human visual cognition system. Our 2-tier intelligent video tracking approach uses a fast, space-based peripheral vision component for holistic video analysis and tunnel vision components for object-based detail analysis, when both are combined, an efficient overall scene analysis framework is produced. The representation for this extracted information must be in a form that is efficient and readily available for upstream applications; including bandwidth limited communications, inference of higher level understanding tasks, and involving other sensor data such as other cameras, aerial imagery, radar imagery etc. The current prototype sensor system uses a wireless mesh-network backbone which can incorporate a range of sensor types including video, acoustic and chemical. The data from these sensors can be correlated and combed for specific targets (vehicles, people etc). The application requirements impose competing richness of context and bandwidth constraints, thus the need for an efficient, flexible and rich representation of the extracted content.