Covariance based Tracking

Tracking with Covariance Descriptors





This work proposes an efficient approximation of a covariance based feature representation for tracking. In contrast to approximated similarity measurements between second order moments, such as Foerstner, we propose to approximate single distributions by specified sampling. We derive an efficient and discriminative feature representation that allows to compute distances between covariance-based descriptors on Euclidean space. This approximated representation perfectly fits to the application of tracking, where efficient similarity measurement significantly controls the efficiency and the real-time capability of the resulting approach. Furthermore, we highlight the advantages of the approximation for learning an object-specific representation during tracking. The experimental evaluation shows results on standard tracking videos and compares our derived approach to state-of-the-art methods based on other covariance representations. The current implementation is running in near real-time performance of currently 6-8 fps in Matlab, using mex-files for feature computation and representation.


Publications:

• Kluckner, S. and Mauthner T. and Bischof, H. (2009): A Covariance Approximation on Euclidean Space for Visual Tracking, Proceedings Annual Workshop of the Austrian Association for Pattern Recognition

Result videos:
Example videos show tracking results on multi-object, non-rigid or partial-occlusion scenes of standard datasets and some additional selected challenging videos.

Download		Hermann Maiers famous accident in Nagano. The Video shows tracking results on a highly dynamic and non-rigid object.	Download		Tracking of an rigid object during partial occlusions. The object is finally lost at the end after a longer total occlusion.
Download		Long time tracking on aerial footage from the VIVID dataset, showing results of scale changes. Results using iPCA updates are shown in by solid lines in comparison to using mean feature update online. Tracking results using iPCA updates are more stable over time and scale changes, due to their object specific subspace representation.	Download		Results of iPCA updates are shown in by solid lines in comparison to using mean feature update. Occlusion are handle directly by the uncertainty measurements of the sub-parts. Hence the tracker handles short-time and partial occlusions. Tracking results using iPCA updates are more stable over time and scale changes, due to their object specific subspace representation.
Download		Results on multi-object and scale scenario in comparison to original method proposed by Porikli et al. Results of our covariance approximation and individual weighting of the parts are shown in by dashed lines in comparison to original covariance features proposed by Porikli et al. Tracking results of our proposed method are more stable over time and scale changes.