Change Detection in Nonlinear Systems and Applications in Shape Analysis

Iowa State University, Ames IA

Investigators

Abstract

ECCS-0725849 Vaswani Change detection is important in most tracking applications, since it is rarely true that the system model is truly time-invariant. Some examples include detecting motion model changes in target tracking or positioning applications; or detecting abnormal shape changes in computer vision/biomedical image analysis applications. In all of the above, the state is not directly observed. The observation is a noise-corrupted and nonlinear function of the state. The effectiveness of particle filters for such nonlinear/non-Gaussian tracking problems is already well known. Often, the changed system model is not known, i.e. the change or abnormality is not characterized. For example, the change may be a gradual one, for a constant velocity target slowly accelerating to a higher speed, or a sudden one. The approach is based on particle filter based algorithms for ``slow"" and ?sudden?, unknown parameter, change detection. Robust design strategies will be developed and tested for realistic applications. Intellectual Merit: Most existing approaches can be classified as either based on adaptive filtering ideas or based on loss-of-track detection. Adaptive filtering approaches are either expensive or unreliable to implement. Loss-of-track based approaches detect abrupt changes almost immediately. However, slow changes, which result in a small loss of track per unit time, usually take a long time to get detected, or sometimes get missed. We propose a novel approach that utilizes the fact that slow changes get partially tracked, and uses this ``tracked part of the change"" for detection using particle filters. Broader Impact: The developed algorithms will impact a large number of positioning, navigation or defense applications that require target tracking; video based surveillance applications that require abnormal behavior detection; biomedical signal/image sequence analysis applications where detected abnormalities can be indicators of disease, as well as many other applications in econometrics, finance, robotics and vision. Educational initiatives will include introduction of a graduate class on Adaptive filtering and Monte Carlo methods; modification of existing undergraduate classes; and senior design projects to implement and compare different shape extraction techniques for various applications.