Intelligent Vision Processor

It is easy to give a robot or other device trying to emulate human ability an ´eye´. Some cameras even exceed the powers of human eyes, but the computations just to measure depth (without interpreting the scene further) are handled by our brain´s extraordinary parallel capabilities and are beyond the ability of any modern single processor. We describe an implementation of a vision system in reconfigurable hardware which measures scene depths at ~ 25 fps in 1 Mpixel images. This implementation uses lookup tables to remove camera distortions and misalignments and Symmetric Dynamic Programming Stereo for matching. It produces depth maps with latencies < 100 scan lines (< 3 ms) for a large misalignment (5deg). Latency can be reduced by precise alignment and lower distortion lenses.