“Stable Multi-Target Tracking in Real-Time Surveillance Video” is a paper that discusses algorithms designed to follow individuals in surveillance video. While many surveillance systems are able to determine the identity of an individual and his general whereabouts, it is harder to do behavioral or biometric analysis without a stable bounding box around him. The processes described in the paper help surveillance systems maintain a more accurate description of the movements of several individuals at once.
After wading through all the jargon and acronyms, I found that this paper is really about using probability models to predict the possible paths an individual might take. Detections are only performed on some of the frames of video. The rest of the frames use interpolation to fill out the path. There is an optimal number of detections, because increasing this number also increases the number of false positives. It is also computationally taxing to do those calculations on every frame.