

A typical commodity camera rarely supports selecting a region of interest to reduce bandwidth, and depending on the extent of image processing, a single CPU may not be sufficient to process data from the camera. Further, such cameras often lack support for synchronized inter-camera image capture, making it difficult to relate images from different cameras. This paper presents a scalable, dedicated parallel camera system for detecting objects in front of a wall-sized, high-resolution, tiled display. The system determines the positions of detected objects, and uses them to interact with applications. Since a single camera can saturate either the bus or CPU, depending on its characteristics and the image processing complexity, the system supports configuring the number of cameras per computer according to bandwidth and processing needs. To minimize image processing latency, the system focuses only on detecting where objects are, rather than what they are, thus reducing the problem's complexity. To overcome the lack of synchronized cameras, short periods of waiting are used. An experimental study using 16 cameras has shown that the system achieves acceptable latency for applications such as 3D games.