Gaussian Mixture Model for Multi-sensor Tracking

We present an algorithm for tracking many objects observed with distributed, non-overlapping sensors. Our method is derived from a proposition that the observations of some constant, intrinsic properties of an object form a cluster (eg. in the color space). However sensors also provide dynamic data about an object like time and location. Tracking is achieved by probabilistic clustering of observations with a Gaussian Mixture Model (GMM). We extend the GMM with auxiliary hidden variables that capture Markov dependencies between points generated by the same kernel. Since the new variables are deterministic given trajectories, the inference in our model can be done effectively with forward-backward propagation. The experiments on real and artificial data suggest that such an approach might be a faster alternative to the existing tracking methods based on MCMC sampling of trajectories.