Autonomous mobile robots, such as the Dante Volcano Explorer developed at CMU (left), place very severe demands on their vision systems.  Robots rely on computer vision algorithms for path planning, obstacle detection and other functions.  These computer vision techniques require a great deal of computing power.  Stereo vision, for example, is often used by robots to produce depth maps of the environment.  Typical stereo vision algorithms perform billions of computational operations for every image processed.  The computational requirements of mobile robots are thus often beyond what can be provided by a PC or workstation class machine.  The Reconfigurable Vision Machine can provide the processing power necessary to address the high computing needs of robotic vision. 
 
Robot computer vision is also time-critical.   For example, an autonomous vehicle moving at high speed and relying on computer vision for obstacle detection must get the data from the vision system in a timely manner.  Vision results which are delayed by more than a tenth of a second are not very useful.  Conventional computer systems frequently do not have the real-time performance or data transfer bandwidth to guarantee vision results at the required rate.  The dedicated, real-time nature of the Reconfigurable Vision Machine makes it an ideal choice in these circumstances. 

The Reconfigurable Vision Machine is currently being used to run the vision system which controls CMU's unmanned, fully autonomous robot helicopter (right). The machine is installed aboard the helicopter and runs algorithms for ground detection, localization, object detection and object tracking.  

Computer vision is also used extensively for industrial control and product inspection.  Often, industrial inspection can be performed by PCs or relatively simple dedicated hardware.  As industrial computer vision has become more common, however, industrial users have become interested in addressing much more difficult inspection tasks.  One example is the inspection of color printing on beverage cans which are printed at 60 cans per second.  This inspection task represents almost a worst-case for vision performance.  The images are high resolution, color, and very fast and this requires very high bandwidth (800*1200 pixels * 3 colors * 60Hz = 164 MB/Sec!) image transfer.  The processing of this 164 MB/Sec data stream is also very involved and demands huge computing power.  Finally, mechanical constraints on the bad can rejection device require the vision system to make a decision on whether to reject a can in a very short time.  This type of industrial application is far beyond the capabilities of workstations, PCs, or off-the-shelf inspection machines.