How do you think the new GigE standards will influence the machine vision industry?
Respond or ask your question now!
The basic data flow model is a generalized form of the model that applies to any computer that is intended to perform vision tasks: on a frame by frame basis the CPU has access to input imagery. Left and right rectified source imagery can be moved into main memory on the PowerPC by direct memory access (DMA). In principle, this data is enough to perform any vision- or stereo-based task. However, the cost of computing stereo depth at frame rate alone would swamp the embedded PowerPC. To perform real-time vision tasks, the PowerPC has access to several other input image sources on each frame: a range image, a foreground image, a "Projection Space" image and some output from a DSP that in turn has access to the source images, range image, foreground image and "projection image."
While not a security application, the problems of an interactive display in New York illustrate how the system worked. Electroland, a Los Angeles-based artistic group that creates multi-disciplinary urban projects, used TYZX PersonTrack for an interactive display for Target on the observation deck of Rockefeller Center.
"The installation was in a breezeway that connects two view terraces," says Damon Seeley, a founding co-partner of Electroland. "The breezeway space has large glass doors on the north and south side, allowing bright sunlight to enter the space during the day. At night, backlighting creates even ambient illumination, but the LED pixels often overwhelm the space with red, green or blue light effects. These widely ranging lighting conditions present huge challenges for 2D computer vision."
The direct measurement of the 3D location of each person creates more robust results than systems based on 2D images alone, TYZX says. Fast frame rates simplify the matching of each person's location from frame to frame. Each stereo camera is calibrated to produce absolute 3D measurements, which simplifies the process of registering the cameras to each other during installation.
Other factors, however, come into play—network bandwidth, power requirements, wiring requirements and physical design. Limited network bandwidth in large installations dictates that most image processing must occur next to the sensor, sending only low bandwidth results such as locations and descriptions of tracked objects, over the network. Each sensor also must host enough processing power to perform tracking in its view space. Each system must be unobtrusive—fans are not practical, for example. This places a limit on the power dissipation. The system also must be small enough to mount easily to the ceiling or wall. Wiring should be minimized; the G2 is designed to run with a single cable, power over Ethernet, or two connections, Ethernet and power. 3D person tracking is a computationally intense task, but the use of hardware accelerated visual primitives offloads most of the processing from the G2 embedded CPU.