Multi-Modal Tracking of Interacting Targets Using Gaussian Approximations

We introduce a modular tracker architecture that combines the advantages of several simple and rapidly performing tracking algorithms. Robust tracking is facilitated by fra merate or near-frame-rate processing. A Kalman filter is used to integrate tracking results across detection modules and over time. Processing regions are smoothly localized by weighting with a Gaussian that is dimensioned according to the target sizes and uncertainties estimated by the Kalman filter. From this Gaussian mask, Gaussian approximations of known nearby targets are subtracted to allow individual tracking of interacting targets, that can be merged and spli t based on Mahalanobis distances and a robust version of connected components. The result is an adaptive tracker that can robustly track at video frame rates several targets , each of which corresponds to one or more individual objects. Its performance degrades gracefully with increased system load.