Powerful, Electo-Optical, Image Processing and Pattern Recognition Device

The recognition and tracking of objects, especially moving objects, in real time from a visual input has been studied extensively but remains a difficult problem. A lot of data must be moved around quickly and many computations are required upon it. We propose a two-part solution, made possible by our novel work on two related product lines. First, to perform feature extraction over the entirety of an image in real time, we propose to develop a high resolution, pin-compatible, 2-D, CCD image sensor with integrated electro-optic, feature-extraction neural network. This small, rugged device will couple a novel, miniature, "distributed optical device" of our design and manufacture to a commercially available image sensor. This modified sensor will plug into the image sensor socket in a video camera and, using the massive parallelism of optics, extract features over the entirety of scenes that are imaged upon it. Second, to increase flexibility and perform neural network and signal processing operations in time and space n real time over the extracted features that are output from the sensor, we propose to develop algorithms that run on our A236 Parallel Signal Processor Module. This small, low cost module provides a unique combination of very high performance, large memory capacity and bandwidth, low power dissipation and very high I/O rates for real time applications.