Scalable vision graph estimation for a vision sensor network

This paper describes a scalable method of estimating a vision graph, in which a pair of camera nodes are connected by an edge if the two nodes share the same field of view, based on local image feature correspondences. The proposed method is implemented in a distributed fashion, meanwhile avoiding the flooding of the image feature information since it can be a bottleneck in achieving scalability. The key idea is to partition the image feature space into a set of disjoint regions so that the correspondence search can be carried out within a partitioned region, with each region served by a different network node independently. Simulated results using real images show that the proposed method achieves reasonable estimation performance while improving the traffic amount and traffic balance greatly.