Tracking across nonoverlapping cameras based on the unsupervised learning of camera link models

A multiple-camera tracking system that tracks the human across cameras with nonoverlapping views is proposed. The system is divided into two phases. In the training stage, the camera link model, including transition time distribution, brightness transfer function, region mapping matrix, region matching weight, and feature fusion weight, is estimated by an unsupervised learning scheme which tolerates well the presence of outliers in the training data. In the testing stage, besides the temporal and holistic color features, region color and region texture features are considered. The systematically integration of multiple cues enables us to perform the effective re-identification. The camera link model keeps being updated during the tracking in order to adapt the change of the environment. The complete system has been tested in a real-world camera network scenarios.