This paper presents a robust method for real-time visual human body tracking. We perform a hierarchical 3-D reconstruction from multiple camera views as a basis for tracking. Blobs attached to a kinematic model are then used to reliably track individual body parts with both volume and color information. We describe how the blob-model is dynamically adjusted to accommodate different body configurations. In tests, our system has proved robust in presence of noisy data and self-occlusions. The system was tested on a single 2GHz PC, with four IEEE 1394 low-resolution webcams. The combined acquisition, reconstruction and tracking ran at 15 fps. Automatic blobs reconfiguration from a naive model to a better one sufficient for most interactive applications. The tracking process proved to be robust to noisy data and self occlusions thanks to the EM statistical framework and the use of color. With a maximal precision of 1.5 cm, our system can capture a broad range of body movements, and the small jitter due to the camera noise can easily be filtered.
© Copyright 2004 Proceedings of the IEEE and ACM International Symposium on MAR. IEEE. Alle Rechte vorbehalten.