Authors:
João F. Henriques
;
Rui Caseiro
and
Jorge Batista
Affiliation:
University of Coimbra, Portugal
Keyword(s):
Visual surveillance, Tracking, real-time, Dynamic Hungarian algorithm, Region covariance matrices.
Related
Ontology
Subjects/Areas/Topics:
Applications
;
Computer Vision, Visualization and Computer Graphics
;
Human-Computer Interaction
;
Methodologies and Methods
;
Model-Based Object Tracking in Image Sequences
;
Motion and Tracking
;
Motion, Tracking and Stereo Vision
;
Pattern Recognition
;
Physiological Computing Systems
;
Real-Time Vision
;
Tracking of People and Surveillance
Abstract:
Tracking is a crucial task in the context of visual surveillance. There are roughly three classes of trackers: the classical greedy algorithms (based on sequential modeling of targets, such as particle filters), Multiple Hypothesis Tracking (MHT) and its variants, and global optimizers (based on optimal matching algorithms from linear programming). We point out the shortcomings of all approaches, and set out to solve the only gaping deficiency of global optimization trackers, which is their inability to work with streamed video, in continual operation. We present an extension to the new Dynamic Hungarian Algorithm that achieves this effect, and show tracking results in such different conditions as the tracking of humans and vehicles, in different scenes, using the same set of parameters for our tracker.