Multicamera fusion for online analysis of structured processes

Abstract

We propose a novel framework for online analysis of visual structured processes, using fusion from multiple cameras. Online recognition is performed through particle filters supported by hidden Markov models. We evaluate three fusion methods, an early fusion, a simple multiplication of the observation probabilities and a multi-stream one implying cross-stream coupling of observations and states. The performance is thoroughly evaluated under two complex visual behavior understanding scenarios: a visual process for table preparation in a kitchen and a real life manufacturing process in an industrial plant. The obtained results are compared and discussed.