Efficient Data Association in Visual Sensor Networks with Missing Detection

Abstract

One of the fundamental requirements for visual surveillance with Visual Sensor Networks (VSN) is the correct association of camera's observations with the tracks of objects under tracking. In this paper, we model the data association in VSN as an inference problem on dynamic Bayesian networks (DBN) and investigate the key problems for efficient data association in case of missing detection. Firstly, to deal with the problem of missing detection, we introduce a set of random variables, namely routine variables, into the DBN model to describe the uncertainty in the path taken by the moving objects and propose the high-order spatio-temporal model based inference algorithm. Secondly, for the problem of computational intractability of exact inference, we derive two approximate inference algorithms by factorizing the belief state based on the marginal and conditional independence assumptions. Thirdly, we incorporate the inference algorithm into EM framework to make the algorithm suitable for the case when object appearance parameters are unknown. Simulation and experimental results demonstrate the effect of the proposed methods.

Distributed data association in smart camera networks using belief propagation

“…By inferring the posterior distribution of the labeling variable based on all appearance and spatio-temporal evidences, the object corresponding to each observation can be determined [Zajdel 2006;Wan and Liu 2011]. However, the computation in association inference is also intractable due to the exponentially explosion of the belief state space.…”

“…However, the computation in association inference is also intractable due to the exponentially explosion of the belief state space. This problem is solved in [Zajdel 2006;Wan and Liu 2011] based on the idea of Assumed Density Filtering proposed in [Boyen and Koller 1998;Boyen et al 1999].…”

“…The algorithms used for comparison in this article include: (i) Centralized EM with appearance based inference, referred as C-app, which is just the standard EM for mixture model [Bilmes et al 1998]; (ii) Centralized EM with forward inference, referred as C-fwd, which realize the E-step with the forward inference described in [Wan and Liu 2011], taking into account both appearance and spatio-temporal information;…”

“…One of the fundamental requirements for achieving these goals is to accurately associate the observations produced by the cameras with the track of each object of interest. This problem is often referred to as data association [Kuo et al 2010;Zajdel 2006;Oh et al 2009;Wan and Liu 2011], or object reacquisition in some literatures [Arth et al 2007]. …”

“…Despite the huge amount of efforts [Gilbert and Bowden 2006;Javed et al 2008;Kuo et al 2010] to overcome theses difficulties, the association performance based solely on appearance is still unsatisfactory. Therefore, spatio-temporal observations, combined with the knowledge of the structure of the monitored region, are always used to improve the accuracy of data association [Pasula et al 1999;Kim et al 2009;Kröse 2005, 2003;van de Camp et al 2009;Zajdel 2006;Oh et al 2009;Goyat et al 2010;Wan and Liu 2011]. However, unlike the appearance observations, spatio-temporal data generated on different cameras are strongly interdependent with each other.…”