Multi-target tracking using dominant sets

Abstract

In this thesis we tried to exploit Dominant set frame work to track multiple people in videos. Since it's unsupervised, given unlabeled patches of the people to follow, we then propose a formulation of the multi-target tracking problem as identifying dominant sets. Dominant set describe very compact structures, which is ideally suites to represent the appearance of a given person in any number of frames as a one cluster. A dominant set is a form of maximal clique that can be applied to edge weighted graphs so that the affinity between all the nodes that are in the set is higher than those which are external to it. We used peeling off strategy to our work, which help us identify all dominant sets in a graph. The application is able to work in both pre-recorded video streams (off-line) and live streaming video (online). Here the data points are the detected persons (patches) in each frame. As we all know, Videos are composed of frames and in each frame there are peoples to be tracked. And we used HOG (histogram of oriented gradient) people detectors to extract the patches. Then each of the detected patches will be treated as a graph node and there will be a similarity comparison between the nodes. In order to capture the individual similarities in people (patches) of similar target and differentiate between different targets, it is compulsory that the graph is made using meaningful and robust similarity measure. We tend to describe people patches with covariance matrix feature descriptors and we build the similarity matrix using distance among covariance in Riemannian manifolds. We finally performed an experiment on different video datasets and got promising good results.