Detecting of Hidden Patterns in Air/Maritime/Human Situation Picture

Dr. Uri Degen (speaker), Dr. Leonid Shvartser, Igor Koro- tayev, Rima Nageris-Gandlin, Alexander Knafel
Israеl, TSG IT Advanced Systems Ltd
This paper deals with various aspects of the pattern recognition in a Situation Picture generated elsewhere. “Situation picture” is the transformation from the actual situation via the “measurement space” containing detections of sensors, possibly multiple and heterogeneous, to the “estimation space” intended to contain the best approximation to the actual situation.
Automatic detection of hidden patterns in the area of interest and consequent alerting facilitate the work of human operator responsible for this area. We present statistical algorithms for real-time recognition in the “situation picture” of various geographical, dynamic and electronic patterns.
In particular, it may be useful to identify simultaneous similar behavior of a group of tracked objects. For example, gathering of people into groups and further joint motion to an important object may be an evidence of criminal or hostile intents. We developed methods of group target identification and tracking in multiple targets area, including group initialization as a set of clusters, group confirmation, group tracking, split and merge groups or a set of groups. Results are illustrated on aircraft and human tracking examples.
Another aspect of trajectory patterns recognition is detection of unusual behavior. Usual behavior of moving objects is defined by entrances, exits, way points, kinematic constraints and other factors. Generally speaking, it is not chaotic and thus can be modeled. Trajectories found to deviate from the routine model are labeled as unusual (abnormal). Unusual trajectories may indicate an abnormal activity. The routine behavior can be described in different ways: as Gaussian Mixtures Model (GMM) on a set of kinematic features for trajectories passing a cell in a scene, as a set of trajectory clusters united on base of Dynamic Warp Distance, as a single class Support Vector Machine on a set of trajectories or as inter-cell probability distribution. Unusual behavior detection is done for every model separately and accomplished with voting. Two different types of trajectories with different kinematics are investigated: ships and humans. It is shown that the features describing fast kinematics are different from the features describing slow kinematics.