A probabilistic framework for combining tracking algorithms

Abstract

For the past few years researches have been investigating enhancing tracking performance by combining several different tracking algorithms. We propose an analytically justified, probabilistic framework to combine multiple tracking algorithms. The separate tracking algorithms considered output a probability distribution function of the tracked state, sequentially for each image. The algorithms may output either an explicit probability distribution function, or a sample-set of it via CONDENSATION. The proposed framework is general and allows the combination of any set of separate tracking algorithms of this kind, even on different state spaces of different dimensionality, under a few reasonable assumptions. The combination may consist of different tracking algorithms that track a common object, as well as algorithms that track separate, albeit related objects, thus improving the tracking performance of each object. In many of the investigated settings, our approach allows us to treat the separate tracking algorithms as "closed boxes". In other words, only the state distributions in the input and output are needed for the combination process. The suggested framework was successfully tested using various state spaces and datasets.

A local-motion-based probabilistic model for visual tracking

“…Since the proposed local-motion model can help resolve ambiguities associated with multiple visually similar targets, it can be used in existing probabilistic multi-cue integration frameworks like [32,8,16], or as extension to multipletarget tracking schemes, such as [31], to increase their robustness when tracking visually-similar targets. Note also that the local-motion-based feature is general enough to be used not only within the framework of particle filters, but also with non-stochastic methods: For example, the discrimination-based trackers such as the recently proposed AdaBoost tracker [33] or the level-set-based blob trackers like [34,35].…”

Low computation and low latency algorithms for distributed sensor network initialization

“…Using mathematical induction, we show that it is possible to sequentially generate particles and weights according to (13) and (14) respectively to represent the cumulative posterior distribution (12).…”

“…The output of our initialization algorithm can be used to initialize various distributed joint tracking (DJT) algorithms such as those described in [11], [12].…”

Thermo-visual feature fusion for object tracking using multiple spatiogram trackers

“…Combining the two modalities would seem to be a rational approach to obtaining optimum system performance and robustness. In relation to object tracking in particular, it is generally recognised that tracking robustness cannot be obtained with one modality (or feature) alone, and that increased robustness can be obtained by combining multiple modalities (also known as multi-cue tracking) in such a way that they can, together, compensate for their individual weaknesses [26,16].…”