Efficient Detection and Mutli-Target Tracking of Persons in Real-Time Video Surveillance
Locating and identifying multiple individuals in a scene are challenging tasks in real-time video surveillance applications. Although tracking allows to locate a person over time, automatically tracking multiple targets under real-world conditions is a challenging problem due to changes in appearance, occlusions and complex backgrounds. Target models are typically adapted for robust discriminative tracking, although representative training samples must be selected on-line such that knowledge corruption is avoided. This project seeks to develop adaptive systems that can robustly detect and track a variable number of people captured in video sequences for real-time visualization and person re-identification applications. Multi-target tracking methods will autonomously create, associate and remove person tracks based on information extracted from the appearance of each person’s head. For efficiency, different heads being tracked will be associated from frame to frame over constrained target search regions. Tracking-by-detection approaches will be developed for robust tracking, where head appearances are modeled using classifiers that are continuously adapted to changes in the operational domain though on-line and incremental learning. In particular, the proposed systems will incorporate accurate head detection using adaptive ensembles of classifiers based on diverse head representations.