Decentralized human tracking in visual sensor networks: using sparse representation for efficient communication

Official URL: http://dx.doi.org/10.1007/978-3-319-10807-0_3

The recent advances in camera sensors and development of new distributed processing algorithms have enabled a new kind of wireless sensor networks namely the visual sensor networks (VSNs). VSNs consist of a network of image sensors, embedded processors, and wireless transceivers which are powered by batteries. The constraints on energy and bandwidth resources challenge setting up a tracking system in VSNs. In this chapter, we present a sparsity-driven decentralized framework for multi-camera human tracking in VSNs. The traditional centralized approaches involve sending compressed images to a central processing unit, which, in the case of severe bandwidth constraints, can hurt the performance of further processing (i.e., tracking) because of low-quality images. Instead, we propose a decentralized tracking framework in which each camera node performs feature extraction and obtains likelihood functions. We propose a sparsity-driven method that can obtain bandwidth-efficient representation of likelihoods. Our approach involves the design of special overscomplete dictionaries that match the structure of the likelihoods and the transmission of likelihood information in the network through sparse representation in such dictionaries. By exploiting information from the sparse representation obtained in the previous frame, we spatially constrain the set of allowed dictionary coefficients in the current frame to reduce the size of the optimization problem and hence, the computation time. Experimental results show that our sparse representation framework is an effective approach that can be used together with any probabilistic tracker and that can provide major savings in communication bandwidth without significant degradation in tracking performance.