Probabilistic data association with the inclusion of attribute information
D.M. Drumheller, Kwang Y. Lee, Leon H. Sibul · 1989
When tracking a target in a cluttered environment, the origin of any measurement of a target's kinematic state is uncertain. That is, sensors may report several measurements of the target. Several approaches have been taken to solve this problem; however, the probabilistic data association filter (PDAF) is attractive since it can produce a good track without excessive computational and memory requirements. Based on the Kalman filter, the PDAF incorporates all measurements into the estimate of the target state using the posteriori probability of each measurement originating from the target. It is reasonable to expect that any algorithm designed to track a target in a cluttered environment would benefit from any additional data not necessarily originating from target kinematics that allows the algorithm to discriminate between the measurement originating from the target and those due to clutter. Such benefits could be a reduction of track loss, and an improved estimate of the target's kinematic state. Accordingly, this dissertation presents a technique for incorporating non-kinematic measurements into the PDAF. Non-kinematic measurements are referred to as attribute data, and are defined as sensed quantities that identify a particular type of target. Attribute data are available in two forms. The sensors may just provide the measured attributes associated with each kinematic measurement, or they may process all the measured attributes and provide the probability of each possible type of target. The latter form of data is referred to as target type measurement. Knowing the probabilistic relationships between the target types and the states each attribute can occupy, we can use the non-kinematic measurements to determine which type of target is most likely being tracked, and what are its most likely attributes. The algorithm utilizes a tree structure called an net, which is a representation of a network of hierarchically organized probabilistic relationships. The tracking technique presented in this dissertation will consist of an inference net used to calculate the target type and attribute probabilities, and a Kalman filter that will track the target's kinematic states with the aid of probabilities provided by the inference net.