Performance Analysis of Graph-based Track Stitching

Recently graph-based tracking (GBT) algorithms were proposed as a new approach to multiple hypothesis tracking (MHT) because of efficiency issues of existing MHT algorithms. This paper analyzes the data association performance of a particular class of GBT algorithms when applied to track stitching problems. The tracklets or track segments to be stitched are the nodes (vertices) in the association or track graph and the arcs (edges) represent possible associations. The best stitching hypothesis is found by solving an optimization problem similar to that of MHT. When tracklet stitching or association likelihoods satisfy the Markov or path-independence assumption, the optimization problem can be reduced to a maximum weight bipartite matching problem. While several polynomial-time bipartite matching algorithms are available, the stitching performance depends on the validity of the Markov or path-independence assumption. This paper examines the effect of this assumption on the association performance of track stitching problems, through a simple analysis and Monte Carlo simulations using a simple track stitching problem. The analysis shows some potential performance loss using the path-independence approximation, as well as potential gain using the path-dependent track likelihood calculation.