Visual tracking of partially observable targets with suboptimal filtering

A comparative study of four suboptimal tracking algorithms that can cope with missing measurements of low-level visual features during partial occlusion is presented. An approach to identify missing measurements in two-dimensional object tracking is formulated. The comparison starts from a symbolic analysis in Kalman filtering and ends with a performance evaluation in Monte Carlo simulations in which the objects manoeuvre and undergo moderate size variations during occlusion. It is found that the algorithm for estimating unobservables from observables outperforms the others in terms of mean square error, robustness and readiness for implementation.