Model-Based 3D Tracking in Multiple Camera Views

Abstract

We present a method for tracking objects in 3D by comparing views from multiple video cameras against a 3D model. Camera-based object tracking is a vast field that until recently has been limited to the 2D (image) domain. With the falling cost of sensors in recent years, the research community has been shifting its focus to 3D tracking using networks of calibrated cameras. The standard approach has been to run traditional 2D tracking in each video and fuse the results. The algorithm described below tracks objects directly in 3D to give increased precision and to overcome limitations inherent in the standard approach. We demonstrate the effectiveness in tracking construction personnel in video sequences of a construction site mock-up, resulting in a reduction in tracking error over current methods.

3D Ground-Truth Systems for Object/Human Recognition and Tracking

“…Note the hardhat being tracked by the iGPS system (white circle). [14] The localization errors between the multi-camera network and the iGPS ground-truth data are presented in Figure 8. …”