A Multi-View Pedestrian Tracking Method in an Uncalibrated Camera Network

Combining multiple observation views has proven beneficial for pedestrian tracking. In this paper, we present a methodology for tracking pedestrians in an uncalibrated multi-view camera network. Using a set of color and infrared cameras, we can accurately tracking pedestrians for a general scene configuration. We design an algorithmic framework that can be generalized to an arbitrary number of cameras. A novel pedestrian detection algorithm based on Center-symmetric Local Binary Patterns is integrated into the proposed system. In our experiments the common field of view of two neighboring cameras was about 30%. The system improves upon existing systems in the following ways: (1) The system registers partially overlapping camera-views automatically and does not require any manual input. (2) The system reaches the state-of-the-art performance when the common field of view of any two cameras is low and successfully integrates optical and infrared cameras. Our experiments also demonstrate that the proposed architecture is able to provide robust, real-time input to a video surveillance system. Our system was tested in a multi-view, outdoor environment with uncalibrated cameras.