Distributed Image and Video Processing: Directed Graphs, Bayesian Estimation, and Hidden Markov Models

Professor Dan Schonfeld
Tuesday, 21.8.2007, 11:30
Room 1061, EE Meyer Building

Traditional image and video processing algorithms are centralized. They rely on an implementation on a large server for efficient processing. However, with the emergence of large camera networks, we must design distributed algorithms that can scale with the size of the network and number of targets. In this talk, we present a general methodology to the design of distributed image and video processing systems. The premise of our approach to distributed processing is the graphical decomposition of complex dynamical systems. We provide a distributed Bayesian approach to multi-object tracking and multi-camera tracking. Implementation of the proposed approach to distributed Bayesian tracking is achieved using a particle-filtering framework. In this framework, multiple particle filters associated with individual targets and cameras collaborate to obtain the joint Bayesian estimate. We subsequently present a distributed multi-dimensional hidden Markov model (HMM). A complex non-causal multi-dimensional HMM is characterized by multiple distributed causal HMMs. The training and classification algorithms of the causal HMMs are derived by extension of the expectation-maximization (EM), general forward-backward (GFB), and Viterbi algorithms to multi-dimensional systems. We use the proposed HMM for multiple interactive motion trajectory classification and natural-man-made image segmentation.

Back to the index of events