Distributed data association in smart camera network via dual decomposition

- Data association by using appearance and spatio-temporal information in the entire network.
- Data association is formulated as an Integer Programming problem.
- Propose two distributed algorithms, L-DD and Q-DD, using dual decomposition.
- Prove the optimality of the solution of L-DD and Q-DD.
- Comparisons are made with the state of the art methods.

One of the fundamental requirements for pedestrian surveillance using smart camera network is the correct association of each person's observations generated on different camera nodes to the person's track. Recently, distributed data association methods that involve only local information processing on each camera node and mutual information exchanging between neighboring cameras have attracted many research interests due to their superiority in large scale applications. In this paper, we propose a new method that performs global data association in a distributed manner by fusing the appearance and spatio-temporal measurements of objects captured by all camera nodes in the entire network. Specifically, we formulate the data association problem 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 the dual decomposition technique. In our algorithms, the original Integer Programming problem is decomposed into several subproblems, which can be solved locally on each smart camera. Different subproblems reach consensus on their solutions in a rigorous way by adjusting their parameters iteratively based on the projected subgradient optimization. The proposed method is simple and flexible, in that (i) we can incorporate any feature extraction and matching technique into our framework to calculate the similarity between observations, which corresponds to the costs of links in our model, and (ii) we can decompose the original problem in any way as long as the resulting subproblem can be solved independently and efficiently. We show the competitiveness of our method in both accuracy and speed by theoretical analysis and experimental comparison with state-of-the-art algorithms.