TY - THES
AU - Fanaswala, Mustafa H.
PY - 2015
TI - Meta-level pattern analysis for target tracking
KW - Thesis/Dissertation
LA - eng
M3 - Text
AB - Classical target tracking operates on a fast time-scale (order of seconds) during which target dynamics are constrained by the physical laws of motion. This dissertation is motivated by pattern analysis on a meta-level (order of minutes or larger) during which the intent of a target manifests. On such a coarse time-scale, Markovian
models quantifying the physical laws of motion are not useful in detecting anomalous behavior. This is due to the hierarchical nature of plans and goal-oriented targets. In this dissertation, several novel stochastic models are devised to capture long-range dependencies within target trajectories for the joint purpose of classification and enhanced tracking. The behavior of targets on the meta level is captured through both positional features (destination) as well as movement patterns (shape). Such features are used with context-free grammar models and reciprocal Markov models (one dimensional Markov random fields) for modeling spatial trajectories with a known end point. The intent of a target is assumed to be a function of the shape of the trajectory it follows and its intended destination. The stochastic grammar models developed are concerned with trajectory shape classification while the reciprocal Markov models are used for destination prediction. Towards this goal, Bayesian signal processing algorithms with polynomial complexity are also presented. The versatility of such models is illustrated with tracking applications in radar-based and optical surveillance.
N2 - Classical target tracking operates on a fast time-scale (order of seconds) during which target dynamics are constrained by the physical laws of motion. This dissertation is motivated by pattern analysis on a meta-level (order of minutes or larger) during which the intent of a target manifests. On such a coarse time-scale, Markovian
models quantifying the physical laws of motion are not useful in detecting anomalous behavior. This is due to the hierarchical nature of plans and goal-oriented targets. In this dissertation, several novel stochastic models are devised to capture long-range dependencies within target trajectories for the joint purpose of classification and enhanced tracking. The behavior of targets on the meta level is captured through both positional features (destination) as well as movement patterns (shape). Such features are used with context-free grammar models and reciprocal Markov models (one dimensional Markov random fields) for modeling spatial trajectories with a known end point. The intent of a target is assumed to be a function of the shape of the trajectory it follows and its intended destination. The stochastic grammar models developed are concerned with trajectory shape classification while the reciprocal Markov models are used for destination prediction. Towards this goal, Bayesian signal processing algorithms with polynomial complexity are also presented. The versatility of such models is illustrated with tracking applications in radar-based and optical surveillance.
UR - https://open.library.ubc.ca/collections/24/items/1.0166215
ER - End of Reference