Electro-optic sensors are evolving from single-color 128 x 128-pixel arrays to multicolor 1000 x 1000-pixel arrays. This increase in sensor size, however, produces a data rate that many on-board processors cannot handle. To overcome this problem, biological vision systems incorporate a small area of high-resolution fovea within a wide field of view having a lower resolution. They can therefore track targets over a wide field of view at a 30-Hz video rate using off-the-shelf PCs.
Based on this approach, a biological vision system has been delivered to the US Air Force Munitions Directorate (Eglin Air Force Base, FL) by Amherst Systems (Buffalo, NY). The reconfigurable prototype vision system consists of a foveal camera, power supply, and host PC. Using a CMOS 128 x 128 reconfigurable, multiresolution, active pixel sensor developed at the Jet Propulsion Laboratory (Pasadena, CA), the imager features direct pixel access for reading high-resolution subsets of the imager array and on-chip pixel averaging to achieve a wide field of view at lower resolutions. These high-resolution subsets become dynamic windows that can track moving objects.
During operation, the targeting algorithms, video processing, and configuration control all execute on the system's Windows NT-based host PC. This system also houses a PCI-DIO-32HS digital I/O board from National Instruments (Austin, TX) that relays real-time control signals from the PC to the camera for imager configuration control. A PCI-MIO-16E-1 A/D board, also from National Instruments, digitizes and stores the video stream for access by the processor and video display hardware.
Three real-time priority programs run under Windows NT—a detection and tracking algorithm (DTA) module, a control module, and a data-acquisition module. The DTA module processes imagery and generates window-of-interest requests. The control module converts window requests into digital control signals that reconfigure the imager at sensor frame rates. The data-acquisition module collects output from the analog-to-digital converter (A/D) board and sends the requested window to the DTA module. In turn, the DTA module generates the next window request. It also processes incoming imagery and establishes a window of interest. The algorithm uses three levels of resolution: peripheral, perifoveal, and foveal (see figure).
The DTA combines centroid- and correlation-based techniques to detect and track moving targets. In centroid-based mode the algorithm acquires a reliable target template. Then it shifts to correlation-based mode. If correlation-based mode loses track of the target, the algorithm returns to the centroid-based mode to establish another target template. This hybrid-tracking algorithm allows the target position to remain stable in the window of interest.
According to Amherst Systems, the next generation of the system will incorporate a 256 x 256-pixel array, and the A/D sensor conversion output will take place near the chip inside the camera housing, eliminating the need for the A/D board in the host PC. The digital I/O board will also be eliminated since communications with the PC will be through a standard Ethernet interface.
