CONTEXT IN ROBOTIC VISION Control for real-time adaptation

摘要

Nowadays, the computer vision community conducts an effort to produce canny systems able to tackle unconstrained environments. However, the information contained in images is so massive that fast and reliable knowledge extraction is impossible without restricting the range of expected meaningful signals. Inserting a priori knowledge on the operative “context” and adding expectations on object appearances are recognized today as a feasible solution to the problem. This paper attempts to define “context” in robotic vision by introducing a summarizing formalization of previous contributions by multiple authors. Starting from this formalization, we analyze one possible solution to introduce context-dependency in vision: an opportunistic switching strategy that selects the best fitted scenario among a set of pre-compiled configurations. We provide a theoretical framework for “context switching” named Context Commutation, grounded on Bayesian theory. Finally, we describe a sample application of the above ideas to improve video surveillance systems based on background subtraction methods.