Accurate video object tracking using a region-based particle filter
Document typeMaster thesis
Rights accessOpen Access
Usually, in particle filters applied to video tracking, a simple geometrical shape, typically an ellipse, is used in order to bound the object being tracked. Although it is a good tracker, it tends to a bad object representation, as most of the world objects are not simple geometrical shapes. A better way to represent the object is by using a region-based approach, such as the Region Based Particle Filter (RBPF). This method exploits a hierarchical region based representation associated with images to tackle both problems at the same time: tracking and video object segmentation. By means of RBPF the object segmentation is resolved with high accuracy, but new problems arise. The object representation is now based on image partitions instead of pixels. This means that the amount of possible combinations has now decreased, which is computationally good, but an error on the regions taken for the object representation leads to a higher estimation error than methods working at pixel level. On the other hand, if the level of regions detail in the partition is high, the estimation of the object turns to be very noisy, making it hard to accurately propagate the object segmentation. In this thesis we present new tools to the existing RBPF. These tools are focused on increasing the RBPF performance by means of guiding the particles towards a good solution while maintaining a particle filter approach. The concept of hierarchical flow is presented and exploited, a Bayesian estimation is used in order to assign probabilities of being object or background to each region, and the reduction, in an intelligent way, of the solution space , to increase the RBPF robustness while reducing computational effort. Also changes on the already proposed co-clustering in the RBPF approach are proposed. Finally, we present results on the recently presented DAVIS database. This database comprises 50 High Definition video sequences representing several challenging situations. By using this dataset, we compare the RBPF with other state-ofthe- art methods.