Conditional filters for image sequence-based tracking - application to point tracking

Abstract

Abstract-In this paper, a new conditional formulation of classical filtering methods is proposed. This formulation is dedicated to image sequence based tracking. These conditional filters allow solving systems whose measurements and state equation are estimated from the image data. In particular, the model that is considered for point tracking combines a state equation relying on the optical flow constraint and measurements provided by a matching technique. Based on this, two point trackers are derived. The first one is a linear tracker well-suited to image sequences exhibiting global dominant motion. This filter is determined through the use of a new estimator, called the conditional linear minimum variance estimator. The second one is a nonlinear tracker, implemented from a conditional particle filter. It allows tracking of points whose motion may be only locally described. These conditional trackers significantly improve results in some general situations. In particular, they allow dealing with noisy sequences, abrupt changes of trajectories, occlusions and cluttered background.

Enhancing the Point Feature Tracker by Adaptive Modelling of the Feature Support

“…This can be achieved by substituting #8 in rennes1, frame 220 #44 in rennes1, frame 220 the error image pixels e i with a difference between the mean value of the warped feature pixel estimated by (11), and the corresponding reference pixel. The pixels of the feature support can finally be identified by testing for:…”

“…Thus, the obtained instances of the feature support usually do not resemble the part of the window projected from a continuous surface. The second approach makes a more explicit check for the temporally consistent feature window pixels, by analyzing the standard deviation estimated by (11). During the motion of the observer, the pixels belonging to a different continuous surface than the one which is consistently tracked, will refer to different points of the scene.…”

“…A critical notion in both approaches is controlling the learning rate of the Gaussian estimates in (11). A fixed value would not be acceptable, since it would imply obtaining different results for different dynamics of the same motion.…”

“…The well behaved portion of a feature window is termed as feature support, while its robust and adaptive detection is the main objective of the paper. The technique is related to robust estimation of the warp parameters [10,11], but is more suitable for detecting correct feature supports which often contain statistical outliers. Here we do not consider updating the reference [12,13,11] despite its potential for increasing the tracking flexibility, since it offers less precision while requiring more processing power.…”

“…The technique is related to robust estimation of the warp parameters [10,11], but is more suitable for detecting correct feature supports which often contain statistical outliers. Here we do not consider updating the reference [12,13,11] despite its potential for increasing the tracking flexibility, since it offers less precision while requiring more processing power. The related research also includes the cumulative similarity transform [14] which is suitable only for tracking homogeneous regions, and the probabilistic filtering of the feature position [15,13,11], which has been used for handling temporary total occlusions.…”