Statistical Cue Integration for Foveated Wide-Field Surveillance

Reliable detection of human activity is an unsolved problem. The main is that low resolution and the unconstrained nature of realistic environments and human behaviourmakeform cues unreliable. Here we argue that reliabilityinfar- or detectioncanstill be achieved byprobabilistic combinationof multiple weak but complementary visual cues that do not depend on detailed form analysis. To demonstrate, we describe a real-time Bayesian algorithmfor localizing human activity in relatively unconstrainedscenes, using motion,background subtractionand skin colour cues. Fast sampling of scalespace is achieved using integral images and a norm that can handle sparse cues without lossofstatistical We show that theprobabilisticapproachfar a representative logical approach in which skin and background subtraction are combined conjunctively Our method is currently used in apre-attentive human activity generating saccadic targetsfor an attentive foveated vision system that faces over a 130 deg of view, al...

• Type: Conference
• Year: 2005
• Where: CVPR
• Authors: Simon J. D. Prince, James H. Elder, Yuqian Hou, Mikhail Sizintsev, Yevgen Olevskiy
• Keywords: Complementary Visual Cues, Computer Vision, CVPR 2005, Human Activity, Human Behaviourmakeform Cues, Skin Colour Cues, Sparse Cues