Distributed Data Association in Smart Camera Networks via Dual Decomposition
One of the fundamental requirements for visual surveillance using smart camera networks is the correct association of each persons observations generated on different cameras. Recently, distributed data association that involves only local information processing on each camera node and mutual information exchanging between neighboring cameras has attracted many research interests due to its superiority in large scale applications. In this paper, we formulate the problem of data association in smart camera networks as an Integer Programming problem by introducing a set of linking variables, and propose two distributed algorithms, namely L-DD and Q-DD, to solve the Integer Programming problem using dual decomposition technique. In our algorithms, the original IP problem is decomposed into several sub-problems, which can be solved locally and efficiently on each smart camera, and then different sub-problems reach consensus on their solutions in a rigorous way by adjusting their parameters based on projected sub-gradient optimization. The proposed methods are simple and flexible, in that (i) we can incorporate any feature extraction and matching technique into our framework to measure the similarity between two observations, which is used to define the cost of each link, and (ii) we can decompose the original problem in any way as long as the resulting sub-problem can be solved independently on individual camera. We show the competitiveness of our methods in both accuracy and speed by theoretical analysis and experimental comparison with state of the art algorithms on two real data sets collected by camera networks in our campus garden and office building.