Abstract :
Automatic video surveillance in uncontrolled outdoor settings is a very challenging computer vision task. Nearly infinite variability of the environmental factors and the open-ended goals of many surveillance problems conspire to create situations where even the most advanced detection, tracking and recognition algorithms falter. While the common academic response to such challenges is to develop new, more powerful algorithms capable of handling a broader range of conditions with acceptable performance, this course of action is sometimes not appropriate from the industrial-commercial point of view. Sometimes systems must be deployed sooner than would allow for the development cycle of complex new algorithms, and must be more robust than most such algorithms can be expected to be on short notice. Under those circumstances, one may look toward better data quality as one means of improving performance while remaining close to the existing state-of-the-art in algorithmic technology. This is often the motivation for deployment of multimodal surveillance systems in the real world. This presentation constitutes a brief foray into design issues for visible-thermal systems, with emphasis on biometric surveillance.
