A fast, feature-driven stereo depth program.

In this paper we describe a fast, feature-driven program for extracting depth information from stereoscopic sets of digitized TV images. This is achieved by two means: in the simplest case, by statistically correlating variable-sized windows on the basis of visual texture, and in the more complex case by pre-processing the images to extract significant visual features such as corners, and then using these features to control the correlation process. The program runs on the PDP-10 but uses a PDP-11/45 and an SPS-41 Signal Processing Computer as subsidiary processors. The use of the two small, fast machines for the performance of simple but often-repeated computations effects an increase in speed sufficient to allow us to think of using this program as a fast 3-dimensional segmentation method, preparatory to more complex image processing. It is also intended for use in visual feedback tasks involved in hand-eye coordination and automated assembly. The current program is able to calculate the three-dimensional positions of 10 points to within 5 millimeters, using 5 seconds of computation for extracting features, 1 second per image for correlation, and 0.1 second for the depth calculation.