Distributed optimization for global data association in non-overlapping camera networks

One of the fundamental requirements for wide-area visual surveillance using non-overlapping camera networks is the correct association of camera´s observations with the tracks of objects under tracking, which is often referred to as data association problem. Most of the current systems work in a centralized manner in that the observations, or the extracted features, on all cameras need to be transmitted to a central server where some association algorithm is running. However, distributed data association scheme, which involves only local information processing on each camera node and information exchanges between neighboring cameras, is more preferable due to a number of well known reasons such as scalability of the system, robustness against single point of failure, and efficiency in use of communication bandwidth, etc. In this paper, we formulate the data association problem as finding the optimal labeling on a factor graph model, and solve the optimization problem under the framework of Dual Decomposition, which can be implemented in a distributed manner. We show the equivalence of the problem of minimizing the energy of the factor graph and that of finding optimal partition of the set of all observations into several subsets, such that all observations belonging to each subset are believed to come from a single object. And the global solution of the dual relaxation of the energy minimization problem is guaranteed to be found by using Dual Decomposition. The effectiveness of the proposed method is demonstrated by comparison with state of the art data association algorithms on real datasets collected by the ten networked cameras mounted in our campus garden.