The objective in direct visual tracking is to find the parameters of a transformation model that align the reference image of a target object to its current image so that a chosen similarity measure between reference and current images is maximized. Tracking is formulated as a registration problem and each method proposed in the literature can be characterized by three specific components: the image similarity measure, the parametric transformation model and optimization strategy. In this page, you will find a study of the following robust image similarity measures commonly used for direct visual tracking:
* Sum of Squared Differences (SSD)
* Sum of Conditional Variance (SCV)
* Normalized Cross Correlation (NCC)
* Mutual Information (MI)
* Cross Cumulative Residual Entropy (CCRE)
If you use our code, please cite:
- Scandaroli G., Meiland M., Richa R., ''Improving NCC-based Direct Visual Tracking'', European Conference on Computer Vision (ECCV), 2012, Firenze, Italy.
* Technical report - Derivation of all similarity measures
The MATLAB code for all experiments shown in this page are available for download. The code follows exactly the formulation described in the technical reports above.
* Package 1 - Sum of Squared Differences (SSD) - (version 1.00 - 12/01/2011)
* Package 2 - Sum of Conditional Variance (SCV) - (version 1.00 - 12/01/2011)
* Package 3 - Normalized Cross Correlation (NCC) - (version 1.00 - 12/01/2011)
* Package 4 - Mutual Information (MI) - (version 1.00 - 12/01/2011)
* Package 5 - Cross Cumulative Residual Entropy (CCRE) - (version 1.00 - 12/01/2011)
If you would like to report a bug or if have any suggestions or comments, please send me an e-mail.
The authors would like to thank Dr. Amaury Dame, Prof. Mark Pickering and Dr. Muhit Abdullah Al for the assistance and helpful comments. This work was partially supported by NIH Bioengineering Research Partnership grant NIH 1R01 EB007969.
Click here to return to my main website.