Probabilistic data association methods in visual tracking of groups

Data association is a fundamental problem when tracking large numbers of moving targets. Most commonly employed methods of data association such as the JPDA estimator are combinatorial and therefore do not scale well to large numbers of targets. However, in many cases large numbers of targets form natural groups which can be efficiently tracked. We describe a method for defining groups based on the position and velocity of targets. This definition introduces a natural set of merging and splitting rules that are embedded into a Kalman filtering framework for tracking multiple groups. In cases where groups of different velocities cross, a general methodology for matching measurements to groups is introduced. This algorithm is based on a modified version of the PDA estimator. It is well suited to handle a high number of measurements and extends naturally to additional grouping constraints such as color or shape.