The problem of multi-target tracking of deforming objects in video sequences arises in many situations in image processing and com- puter vision. Many algorithms based on finite dimensional particle fil- ters have been proposed. Recently, particle filters for infinite dimensional Shape Spaces have been proposed although predictions are restricted to a low dimensional subspace. We try to extend this approach using pre- dictions in the whole shape space based on a Sobolev-type metric for curves which allows unrestricted infinite dimensional deformations. For the measurement model, we utilize contours which locally minimize a segmentation energy function and focus on the multiple contour track- ing framework when there are many local minima of the segmentation energy to be detected. The method detects figures moving without the need of initialization and without the need for prior shape knowledge of the objects tracked.
|Titolo:||Multiple Object Tracking via Prediction and Filtering with a Sobolev-Type Metric on Curves|
|Titolo del libro:||Computer Vision -- ECCV 2012. Workshops and Demonstrations|
|Data di pubblicazione:||2012|
|Nome del convegno:||European Conference on Computer Vision eccv 2012|
|Appare nelle tipologie:||4.1 Contributo in Atti di convegno|