Using group knowledge to track multiple vehicles moving across terrain
Maria L. Gini, Edward Sobiesk · 2000
Multitarget tracking is an essential process for any surveillance system utilizing sensors and computer systems to interpret an environment. A multitarget terrain-based tracking system attempts to maintain an accurate description of an environment in which ground-based vehicles are operating. It does this through the fusion of intermittent sensor reports from the environment with a priori knowledge. This knowledge includes both the characteristics of the environment and the characteristics of the entities operating in the environment. When multiple military vehicles are moving across terrain, most vehicles will be moving as part of a group instead of autonomously. Vehicles that are located within a certain number of meters of each other and are moving in the same general direction with an assumed unity of purpose are considered a group. A group of military vehicles has characteristics that can add additional information to a system. This dissertation creates a clustering algorithm to identify the groups within a vehicular data set and then extracts some of their characteristics. These characteristics are then used as knowledge and integrated into the estimation process. A series of experiments using three classic state estimation models and large, real world data sets shows that the proposed methodology significantly improves state estimation for ground-based vehicles moving across terrain. The major contributions of this dissertation include: (1) demonstrating experimentally using real data that group knowledge improves state estimation for ground-based vehicles moving across terrain, (2) an algorithm for unsupervised identification of groups within a multitarget domain, (3) methods for extracting knowledge from groups, (4) a method for integrating group knowledge into state estimation models, and (5) a paradigm for the use of group knowledge in modeling that applies specifically to the tracking domain and more generally to the domain of motion modeling.