The multiple hypotheses tracker (mht) is recognized as an optimal tracking method due to the enumeration
of all possible measurement-to-track associations, which does not involve any approximation in its original
formulation. However, its practical implementation is limited by the NP-hard nature of this enumeration. As
a result, a number of maintenance techniques such as pruning and merging have been proposed to bound the
computational complexity. It is possible to improve the performance of a tracker, mht or not, using feature
information (e.g., signal strength, size, type) in addition to kinematic data. However, in most tracking systems,
the extraction of features from the raw sensor data is typically independent of the subsequent association and
filtering stages. In this paper, a new approach, called the Judicious Multi Hypotheses Tracker (jmht), whereby
there is an interaction between feature extraction and the mht, is presented. The measure of the quality of feature
extraction is input into measurement-to-track association while the prediction step feeds back the parameters
to be used in the next round of feature extraction. The motivation for this forward and backward interaction
between feature extraction and tracking is to improve the performance in both steps. This approach allows for
a more rational partitioning of the feature space and removes unlikely features from the assignment problem.
Simulation results demonstrate the benefits of the proposed approach.