DSST Object Tracking
The Discriminative Scale Space Tracker (DSST), proposed in [2], extends the Minimum Output Sum of Squared Errors (MOSSE) tracker [1] with robust scale estimation. The MOSSE tracker works by training a discriminative correlation filter on a set of observed sample gray scale patches. This correlation filter is then applied to estimate the target translation in the next frame. The DSST additionally learns a one-dimensional discriminative scale filter, that is used to estimate the target size. The scale filter is trained by extracting several sample patches at different scales around the current target position in the image. Each sample is represented by a fixed-length feature vector based on HOG. These samples are used to learn a multi-channel one-dimensional discriminative filter for scale estimation. This scale filter is generic and can be combined with any tracker that is limited to only estimating the target translation. Given a new image, the DSST first applies a translation filter to obtain the most probable target location. The scale filter is then applied at this location to estimate the target size. The tracking model is then updated with the new information information of the target and background appearance. For the translation filter, we combine the intensity features employed in the MOSSE tracker with a pixel-dense representation of HOG-features.
[1] D. S. Bolme, J. R. Beveridge, B. A. Draper, and Y. M. Lui. Visual object tracking using adaptive correlation filters. In CVPR, 2010.
[2] M. Danelljan, G. Häger, F. Shahbaz Khan, and M. Felsberg. Accurate scale estimation for robust visual tracking. In Proceedings of the British Machine Vision Conference (BMVC), 2014.
Check the demo video as follows:
[1] D. S. Bolme, J. R. Beveridge, B. A. Draper, and Y. M. Lui. Visual object tracking using adaptive correlation filters. In CVPR, 2010.
[2] M. Danelljan, G. Häger, F. Shahbaz Khan, and M. Felsberg. Accurate scale estimation for robust visual tracking. In Proceedings of the British Machine Vision Conference (BMVC), 2014.
Check the demo video as follows:
The experiment result is as follows: