Continuous media and machine vision distributed, active stereo reconstruction

This paper presents the framework and initial results from a distributed computer vision system to perform active stereo vision. We solve tasks in computer vision using a multi-level, distributed continuous media environment. At the highest level, the system recovers complete 3-D models. The intermediate level is a distributed algorithm for reconstructing multiple frames of the input sequence, operating on a high-speed ATM network of workstations and using a distributed, shared-memory model to store results. Control of camera parameters such as focus at the lowest level uses cues computed directly in the compressed image stream. The image stream is compressed by inexpensive multimedia boards, and we exploit the resulting compressed stream by obtaining low-level control information without complete decompression.