Measurement-to-track cost assignment computation when tracking with the IMM estimator

In the multi-target tracking problem, the association of measurements to tracks is typically conducted using an m times n cost matrix, where there are m tracks and n measurements. Each element of this cost matrix is computed based on some measure of the statistical distance between the estimated track state and the measurement. A common approach is to calculate a negative log-likelihood value that includes the uncertainty of the estimate and measurement. When tracking with an interacting multiple model (IMM) estimator, multiple mode likelihood values are possible for each measurement/track pair. Computing the cost assignment matrix given these likelihoods is a critical design issue. This paper compares two approaches for computing the cost assignment matrix. The first computes the likelihood with the mode that maximizes the likelihood for each given measurement. An alternate approach evaluates the pure likelihood of the Gaussian mixture. To focus on data association, we assume that the objects are perfectly resolved. The impact of imperfect object resolution is reserved for future work