Adaptive Regression Tracking
This project is concerned with developing fast regression based tracking algorithms, that adapt to target appearance changes online, and remove the need for hard-coded models and an offline learning stage.
This work proposes an approach to tracking by regression that uses no hard-coded models and no offline learning stage. The Linear Predictor (LP) tracker has been shown to be highly computationally efficient, resulting in fast tracking. Regression tracking techniques tend to require offline learning to learn suitable regression functions. This work removes the need for offline learning and therefore greatly increases the applicability of the technique. The online-LP tracker can simply be seeded with an initial target location, akin to the ubiquitous Lucas-Kanade algorithm that tracks by registering an image template via function optimisation.
A fundamental issue for all trackers is the representation of the target appearance and how this representation is able to adapt to changes in target appearance over time. The two proposed methods, LP-SMAT and LP-MED, demonstrate the ability to adapt to large appearance variations by incrementally building an appearance model that identifies modes or aspects of the target appearance and associates these aspects to the Linear Predictor trackers to which they are best suited. Experiments comparing and evaluating regression and registration techniques are presented along with performance evaluations favourably comparing the proposed tracker and appearance model learning methods to other state of the art simultaneous modelling and tracking approaches.
Available online from publisher http://www.springerlink.com/content/B302N43642876Q85
International Journal of Computer Vision, IJCV , 2011
British Machine Vision Conference, BMVC 2008, Vol 1, pp33-43, Sept 2008
Camera-shake: Tracking results for a video sequence containing considerable shaky camera motion. The solid red box shows the LP-MED tracker result. The solid yellow box shows the Online-Boost tracker result. The dashed yellow box shows the Semi-Online-Boost tracker result. All trackers are initialised in the first frame with no offline learning and no prior models.
Car-Surveillance: Tracking results for a typical surveillance sequence (PETS'2000). The red solid box shows the LP-MED tracker. The yellow dashed box shows the Online-Boost tracker. All trackers are initialised in the first frame with no offline learning and no prior models.
Tracking results for the Dudek sequence. Tracker key: Dark blue - LK, black - LP-FLOCK, green - LK-SMAT, light blue - LP-SMAT, red - LP-MED, yellow - Online-Boost.
Runner: Tracking results for a typical athletics sequence. Tracker key: Dark blue - LK, black - LP-FLOCK, green - LK-SMAT, light blue - LP-SMAT, red - LP-MED, yellow - Online-Boost.
Sequences and ground truth - 314MB
Links to code:
- Code for the LK and LK-SMAT tracker, are available from the following link. A number of options are available (6 similarity functions including Sum of Square Differences and Partial Volume Estimated MI, 4 warp functions including Translation and Affine), and 5 optimisation methods including Simplex and Levenberg-Marquardt):
- Code for the Online boosting trackers used for comparison is available from:
Last updated: 2014-03-14