A pixel for measuring two-dimensional (2-D) visual motion with two one-dimensional (1-D) detectors has been implemented in very large scale integration. Based on the spatiotemporal feature extraction model of Adelson and Bergen, the pixel is realized using a general-purpose analog neural computer and a silicon retina. Because the neural computer only offers sum-and-threshold neurons, the Adelson and Bergen's model is modified. The quadratic nonlinearity is replaced with a full-wave rectification, while the contrast normalization is replaced with edge detection and thresholding. Motion is extracted in two dimensions by using two 1-D detectors with spatial smoothing orthogonal to the direction of motion. Analysis shows that our pixel, although it has some limitations, has much lower hardware complexity compared to the full 2-D model. It also produces more accurate results and has a reduced aperture problem compared to the two 1-D model with no smoothing. Real-time velocity is represented as a distribution of activity of the 18 X and 18 Y velocity-tuned neural filters
Date of this Version
Vision chips, visual motion detection, VLSI neural filters, Adelson and Bergen
Date Posted: 21 October 2005
This document has been peer reviewed.