Multi-camera people tracking using Bayesian networks

Abstract

A multi-camera tracking system is considered for video surveillance, where the cameras have non-overlapping views and the system is required to be robust under different lighting conditions. As people enter the scene, they are segmented out as foreground objects. Facial features, texture and color features of clothing, and likely location of people are extracted and matched with like data stored in a database that is dynamically constructed. The top 3 matches are used in a Bayesian Network, together with a face detection confidence measure, and the likelihood of a person's location for inferencing. To establish identity, results over 16 consecutive frames are used in a majority voting scheme. When the person is identified, the extracted data is stored in the database for future use. To assess the performance of the system, experiments were conducted on a database of 11 people, simulating 4 different camera views. Identification accuracies in excess of 95.45% were obtained in different experiments.