Set Theory Correlation Free Algorithm for HRRR Target Tracking

One challenge of simultaneous tracking and identification of targets is the fusion of continuous and discrete information. Recently a few fusionists including Mahler [1] and Mori [2] are using a set theory approach for a unified data fusion theory which is a correlation free paradigm [3]. This paper uses the set theory approach as a basis for a method of fusing kinematic-continuous data and identification-discrete feature information. The set of features are high range resolution radar range-bin locations and amplitudes which are collected over a small aperture and a scrambled method is used to order a feature set. Once features are ordered, a recursive belief filter operates in feature space to combine track and identification measurements. The intersection of track and identification methods results in a simultaneous tracking and identification algorithm which accumulates evidence for belief in targets and rules out non-plausible targets 1.0 Introduction Multitarget tracking in the presence of clutter has been investigated through the use of data association algorithms [4] such as the joint-probability data association (JPDAF). Likewise, other multisensor fusion algorithms have focused on tracking targets from multiple look sequences such as the multiresolution wavelet-based approach formulated by Hong [5]. One inherent limitation of current algorithms is that the information used to track targets is based only on kinematic measurements. Recently, algorithms have been proposed for feature-aided tracking. Feature-aided tracking uses object features to help discern targets in the presence of clutter such as high-range resolution radar(HRRR) [6]. Typically an image analyst is required to abstract target type data to update tracks. Once the human has a belief or a hypothesis in the target type, the tracking algorithm can be updated. Multiple Hypothesis Tracking (MHT), developed by Reid [7] and based on a multiple hypothesis estimation (MHE) [4], might expand the set of hypothesis to an unmanagble number. In order to implement the algorithm, hypothesis management is typically performed. Using target feature-set information reduces the number of target hypotheses. Three HRR tracking and classification algorithms have been proposed. In 1996, Stone [8] proposed a non-linear Bayesian likelihood ratio tracker (LRT) with an evidential accrual and data fusion approach to control the number of targets and compared it to the MHT algorithm. Following this work, Metron, applied the approach to shipboard HRR [9]. For the second method, Jacobs and O’Sullivan added tracking to their Bayesian HRR ATR algorithm and computed joint likelihood probabilities [10] with applications. Kastella has adapted the work of Jacobs and uses scatter-centering models for a nonlinear joint tracking and recognition algorithm based on joint probability density functions, but much of his work is simulated [11]. Kastella and Musick are using a joint-multiprobability algorithm, or JMP, to associate classifications and track updates. The third algorithm is that of Layne [12,13] an automatic target recognition and tracking filter (ATRF) in a multiple model estimator (MME) approach for HRRR signatures. Layne’s work can be considered an extension of the MME from Libby and Maybeck [14], who simulated an HRRR tracker. These approaches, although influential in this work, rely on the Bayes’ rule for identification where the most probable target is selected. A limitation of using a Bayesian analysis is that it does not capture incomplete knowledge. For instance, there are times when unknown targets might be of interest that are not known at algorithm initiation. At other times, there are unknown number of targets to track or targets not trained for classification. We seek to expand on these tracking and ID algorithms for HRRR signatures, by allowing for the capability to discern unknown relevant targets and reject non-plausible targets. The method used to augment track-association uncertainties is a set-theory approach which helps enhance track quality. Many tracking algorithms utilize information from multiple sources that measure the target at a single resolutional level. Results have proven well for distributed filtering of multiresolutional signals [15,16]. The ability to process Approved for public release; distribution is unlimited.