Distributed Vision Networks for Human Pose Analysis

Multi-camera networks offer potentials for a variety of novel human-centric applications through provisioning of rich visual information. Local processing of acquired video at the source camera facilitates operation of scalable vision networks by avoiding transfer of raw images. Additional motivation for distributed processing stems from an effort to preserve privacy of the network users while offering services in applications such as assisted living. Yet another benefit of processing the images at the source is the flexibility it offers on the type of features and the level of data exchange between the cameras in a collaborative processing framework. In such a framework data fusion can occur across the three dimensions of 3D space (multiple views), time, and feature levels.

In this chapter collaborative processing and data fusion mechanisms are examined in the context of a human pose estimation framework. For efficient collaboration between the cameras under a low-bandwidth communication constraint, only concise descriptions of extracted features instead of raw images are communicated. A 3D human body model is employed as the convergence point of the spatiotemporal and feature fusion. The model also serves as a bridge between the vision network and the high-level reasoning module, which can extract gestures and interpret them against the user's context and behavior models to arrive at system-level decisions. The human body model also allows the cameras to interact with one another on the initialization of feature extraction parameters, or to evaluate the relative value of their derived features.