Segmentation Based Particle Filtering for Real-Time 2D Object Tracking

Abstract

We address the problem of visual tracking of arbitrary objects that undergo significant scale and appearance changes. The classical tracking methods rely on the bounding box surrounding the target object. Regardless of the tracking approach, the use of bounding box quite often introduces background information. This information propagates in time and its accumulation quite often results in drift and tracking failure. This is particularly the case with the particle filtering approach that is often used for visual tracking. However, it always uses a bounding box around the object to compute features of the particle samples. Since this causes the drift, we propose to use segmentation for sampling. Relying on segmentation and computing the colour and gradient orientation histograms from these segmented particle samples allows the tracker to easily adapt to the object’s deformations, occlusions, orientation, scale and appearance changes. We propose two particle sampling strategies based on segmentation. In the first, segmentation is done for every propagated particle sample, while in the second only the strongest particle sample is segmented. Depending on this decision there is obviously a trade-off between speed and performance.

We perform an exhaustive quantitative evaluation on a number of challenging sequences and compare our method with the number of state-of-the-art methods previously evaluated on those sequences. The results we obtain outperform majority of the related work, both in terms of the performance and speed.

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.