Graph approaches for data association

The main problem in multiple object tracking is data association, which has a natural representation as a graph. This paper reviews two different graph approaches for solving the data association problem. The first approach starts with a track graph where the nodes are sensor reports and the edges are possible associations between sensor reports. Solution of the general problem requires combinatorial generation of tracks and optimization by integer linear programming or multidimensional assignment. When the likelihoods satisfy a Markov property, e.g., in track stitching, explicit track generation is not needed and efficient polynomial time algorithms such as bipartite matching or minimum cost network flow can be used. The second and more recent approach represents the joint probability distribution of the association variables and other random variables by a probabilistic graphical model. Distributed inference techniques such as message passing are then used to find the probabilities of associations or the best association hypothesis.