Tracking intelligent objects in terrain

Target-tracking techniques typically rely on the Kalman filter algorithm to predict the motion of moving objects based solely on past observations of their trajectory. While such approaches work well when the objects of interest are subatomic particles, aircraft, or projectiles, they fail to account for the movement of intelligent objects in terrain. In the problem of tracking ground objects, terrain features (e.g., roads and obstacles) induce frequent accelerations and maneuvers, thereby greatly reducing the significance of the past trajectory for predicting the future trajectory of objects. Yet, the knowledge of underlying terrain features and of information pertaining to object behavior makes many maneuvers predictable. The Kalman filter algorithm, however, does not allow the integration of constraints and is therefore not appropriate for terrain-based tracking. The approach presented in this thesis is based upon a motion model propagating the locational probability densities of objects over a discrete state-space. Whereas some authors have approached the problem by using local terrain information in their motion model, the originality of this research consists in (i) the use of path planning methods (in particular potential function methods) for modeling object motion based upon remote terrain features and (ii) the development of an adaptive method capable of updating the most relevant parameters describing object motion to account for the latest observations of the trajectories. These parameters include: (i) attractiveness of candidate goal regions, and (ii) factors determining the attitude of targets towards roads. While their performance was largely dependent upon the quality of the terrain and target information provided by the user, the terrain-based trackers always outperformed the Kalman filter in simulations based on real data.