Distributed EM Learning for Appearance Based Multi-Camera Tracking

Abstract

Visual surveillance in wide areas (e.g. airports) relies on cameras that observe non-overlapping scenes. Multi-person tracking requires re-identification of a person when he/she leaves one field of view, and later appears at another. For this, we use appearance cues. Under the assumption that all observations of a single person are Gaussian distributed, the observation model in our approach consists of a Mixture of Gaussians. In this paper we propose a distributed approach for learning this MoG, where every camera learns from both its own observations and communication with other cameras. We present the Multi-Observations Newscast EM algorithm for this, which is an adjusted version of the recently developed Newscast EM. The presented algorithm is tested on artificial generated data and on a collection of real-world observations gathered by a system of cameras in an office building.

Distributed data association in smart camera networks using belief propagation

“…(iii) Centralized EM with the forward-backward inference in Wan and Liu [2011], referred as C-bwd; (iv) Centralized EM with MCMC inference [Kim et al 2009], referred as C-MCMC, some modification in the trajectory posterior evaluation has been made to make the original algorithm suitable in case of missing detection [Wan and Liu 2011]; (v) Distributed EM with appearance based inference, referred as D-app, which is the algorithm described in [Mensink et al 2007]. However, we find in simulation that when the number of observations is small, the performance of the algorithm in [Mensink et al 2007] is unacceptable.…”

“…Nowak [2003] calculates the global sufficient statistics required in M-step by accumulating the local sufficient statistics on each sensor nodes in a prescribed path through the networks. Mensink et al [2007] proposed a multiobservations extension of the newscast EM algorithm [Kowalczyk and Vlassis 2005] where the averages in M-step are calculated using gossip-based protocol. Gu [2008] proposed a distributed EM over sensor networks where a consensus filter [Olfati-Saber and Shamma 2005] is used to diffuse local sufficient statistics to neighbors and estimate global sufficient statistics in each nodes.…”

Agentification of Markov model-based segmentation: Application to magnetic resonance brain scans

“…Nowak [17] proposed a distributed EM algorithm for density estimation in sensor networks. Mensink et al [18] proposed Multi-Observation Newscast EM for multi-camera tracking where every camera learns from both its own observations and communication with other cameras. These approaches are envisaged in a data fusion context: the algorithm is distributed but only one model is considered.…”

A Survey of Distributed Computer Vision Algorithms

“…The idea seems promising and a more fully distributed system would be interesting. Mensink et al (Mensink, Zajdel, and Krose, 2007) described a multi-camera tracking algorithm in which the observation at each camera was modeled as a mixture of Gaussians, with each Gaussian representing the appearance features of a single person. The authors extended the gossip-based Newscast EM algorithm (Kowalczyk and Vlassis, 2004) discussed in Section 2 to incorporate multiple observations into the iterative, distributed computation of means, covariances, and mixing weights in the Gaussian mixture model.…”