3D shape-encoded particle filter for object tracking and its application to human body tracking
Journal on Image and Video Processing - Anthropocentric Video Analysis: Tools and Applications
| Hi-index | 0.15 |
Consider the problem of detecting and localizing a faint object moving in an 驴essentially stationary驴 background, using a sequence of two-dimensional low-SNR images of the scene. A natural approach consists of 驴digitizing驴 each snapshot into a discrete set of observations, sufficiently (perhaps not exactly) matched to the object in question, then tracking the object using an appropriate stochastic filter. The tracking would be expected to make up for the low signal-to-noise ratio, thus allowing one to 驴coherently驴 process successive images in order to beat down the noise and localize the object. Thus, 驴tracking驴 here does not refer to the ususal notion of detecting then tracking: rather, we track in order to detect. The problem then becomes one of choosing the appropriate image representation as well as the optimal (and necessarily nonlinear) filter. We propose exact and approximate solutions using wavelets and the Zakai equation. The smoothness of the wavelets used is required in the derivation of the evolution equation for the conditional density giving the filter, and their orthogonality makes it possible to carry out actual computations of the Ito- and change-of-gauge-terms in the algorithm effectively.