A particle filtering approach to abnormality detection in nonlinear systems and its application to abnormal activity detection

We study abnormality detection in partially observed nonlinear dynamic systems tracked usingparticle filters. An ‘abnormality’ is defined as a change in the system model, which could be drastic
or gradual, with the parameters of the changed system unknown. If the change is drastic the particle
filter will lose track rapidly and the increase in tracking error can be used to detect the change.
In this paper we propose a new statistic for detecting ‘slow’ changes or abnormalities which do
not cause the particle filter to lose track for a long time. In a previous work, we have proposed a
partially observed nonlinear dynamical system for modeling the configuration dynamics of a group of
interacting point objects and formulated abnormal activity detection as a change detection problem.
We show here results for abnormal activity detection comparing our proposed change detection
strategy with others used in literature.