Multi-Observations Newscast EM for Distributed Multi-Camera Tracking

Visual surveillance in wide areas (e.g. airports) relies on multiple cameras which observe non-overlapping scenes. The focus of this thesis is multi-person tracking, where the task is to maintain a person's identity when he or she leaves the field of view of one camera and later re-appears at another camera. While current wide-area tracking systems are central systems, we propose to use a distributed system; where every camera learns from both its own observations and communication with other cameras. Multi-person tracking can be seen as a data association problem, where the observations of the same person, gathered from different cameras, have to be clustered into trajectories. For this correspondence between a person's identity and an observation, we use appearances features (such as colour or height) and spatial-temporal features (such as the motion of a person from one camera to another). Under the assumption that all appearances of a single person are Gaussian distributed, the appearance model in our approach consists of a Mixture of Gaussians. The Expectation-Maximization algorithm can be used to learn the parameters of the Mixture of Gaussians. In this thesis we introduce Multi-Observations Newscast EM to learn the parameters of the Mixture of Gaussians from distributed observations. Each camera learns its own Mixture of Gaussians model. Multi-Observations Newscast EM uses a gossip-based protocol for the M-step. We provide theoretical evidence, and using experiments show, that in an M-step each camera converges exponentially fast to the correct estimates. We propose to initialize Multi-Observations Newscast EM with a distributed K-Means to improve the performance. We found that Multi-Observations Newscast EM performs equally to a standard EM algorithm. In this thesis we present two probabilistic models for multi-person tracking. The first model uses only appearance features, while the second model also uses spatial-temporal models. Both models are implemented in a central system and in a distributed system. The distributed system uses Multi-Observations Newscast EM, while the central system uses a standard EM. The two models are tested on artificial data and on a collection of real-world observations gathered by several cameras in the university building. The results show that the central system and the distributed system perform equally well. While the more elaborate model, which uses appearance features and spatial-temporal features, outperforms the simple model.