R. Sarpeshkar, W. Bair, and C. Koch. Visual motion computation in analog VLSI using pulses. Neural information processing systems 5, S. Hanson, J. Cowan, C. Giles editors. Morgan Kaufman, San Mateo, pp. 781-788, 1993.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... in computational models for motion estimation (e.g. 2, 21, 37] In the case of analog VLSI the main velocity sensors are the Reichardt Correlation detectors inspired from the fly visual system [27, 66] A number of circuits have been built using this detector or related structures (e.g. [5, 9, 45, 48, 53, 56, 66, 67]) In chapter 3 we present other structures also used in analog VLSI for motion estimation. The signal provided by the motion detectors varies with input contrast, spectral contains, local geometry, etc. Therefore, the output of the motion detectors must be still processed in order to obtain an ....

....frequencies [30, 31] ffl Correlation based: velocity is obtained by combining the output of motion sensors based on the correlation of two delayed neighbor pixels. They are inspired from the visual system of the fly [66] They are the more successful schema for actual motion circuits (e.g. [9, 53, 56, 67]) ffl Time of travel based: velocity is computed by measuring the traveling time of a single feature between two pixels [45, 67] ffl Phase based: velocity is obtained from the phase behavior of band pass filters [26] Some of these algorithms are equivalent and, therefore, will yield similar ....

[Article contains additional citation context not shown here]

R. Sarpeshkar, W. Bair, and C. Koch. Visual motion computation in analog VLSI using pulses. Neural information processing systems 5, S. Hanson, J. Cowan, C. Giles editors. Morgan Kaufman, San Mateo, pp. 781-788, 1993.

....processes within sensors, such as an integrate and fire driven by a continuously varying analog photocurrent. This computational domain may hold the possibility of discovering silicon analogues to the enormously complex computational operations which the biological nervous system performs (see eg [7, 4, 8, 1]. For this reason, much attention has been paid in earlier work to accurate modelling of biological neural tissue. Our starting point in this paper, however, is an exploration of more abstract elementary computational methods in the pulse domain. We propose a set of elementary operations, similar ....

Sarpeshkar R., Bair W., Koch C. "Visual motion computation in analog VLSI using pulses". Neural Information Processing Systems 5 Eds. S Hanson, J Cowan, C Giles. pp 781-788. San Mateo: Morgan Kaufman, 1993.

....dimension or 2x1D by creating a two dimensional array out of one dimensional motion units. The best local information that one can get with these feed forward computation schemes is so called Normal Flow where only local velocities orthogonal to the intensity edge, or gradient, are calculated [3, 2, 4, 5, 11, 7, 1]. The use of such normal flow information to achieve motion segmentation has been shown [8] However, this was done in one dimension where the aperture problem does not occur and feed forward computation of local motion is well defined. In two dimensions this is no longer the case and normal flow ....

R. Sarpeshkar, W. Bair, and C. Koch. Visual motion computation in analog vlsi using pulses. In NIPS, pages 781--788. NIPS, 1993.

.... 2D velocity sensor based on a gradient method [7] showed poor performance, subsequent efforts concentrated on implementing temporal correspondence algorithms, ranging from pure correlation schemes [8] to algorithms performing correlation type motion computation on extracted image tokens [9] 10] [11] and to token based time of travel methods [11] 12] 13] 14] Most of these previous analog VLSI motion sensors either only responded robustly to stimuli of very high contrasts [7] 11] or had an output signal that strongly depended on contrast and or illumination, as well as velocity [8] ....

.... showed poor performance, subsequent efforts concentrated on implementing temporal correspondence algorithms, ranging from pure correlation schemes [8] to algorithms performing correlation type motion computation on extracted image tokens [9] 10] 11] and to token based time of travel methods [11], 12] 13] 14] Most of these previous analog VLSI motion sensors either only responded robustly to stimuli of very high contrasts [7] 11] or had an output signal that strongly depended on contrast and or illumination, as well as velocity [8] 10] 12] Another problem with some ....

[Article contains additional citation context not shown here]

R. Sarpeshkar, W. Bair, and C. Koch, "Visual motion computation in analog VLSI using pulses," Advances in Neural Information Processing Systems 5, D.S. Touretzky, ed., pp. 781-788. San Mateo, Calif.: Morgan Kaufman, 1993.

No context found.

R. Sarpeshkar, W. Bair, and C. Koch, "Visual motion computation in analog VLSI using pulses," in Advances in Neural Information Processing Systems 5, D. S. Touretzky, Ed. San Mateo, CA: Morgan Kaufmann, 1993, pp. 781--788.

....implemented as a hexagonal array of 26 Theta26 pixels [18] compared the velocity of temporal edges traveling across the image plane with the preset velocity of pulses propagating through delay lines. The above mentioned model for motion perception of insects [13] was also applied to image tokens [19]. Using a spatial edge detector, that was constructed with two resistive grids to approximate a difference ofGaussians operator [20] as input stage, it performed a binary correlation of delayed and undelayed voltage pulses from adjacent pixels respectively. While such correlation based algorithms ....

....stage differ significantly among the two motion sensors, while the same edge detector is used for both of them. The facilitate and trigger (FT) motion circuit measures the overlap time of binary pulses of a given width produced by the two adjacent pulse shaping circuits, as shown in Fig. 1(b) [19]. The facilitate and sample (FS) motion circuit samples a time varying voltage pulse generated by one pulse shaping circuit with a thin voltage spike generated by an adjacent pulse shaping circuit, as shown in Fig. 1(c) 6] For each FT or FS motion circuit, one pulse (P 1 ) acts as a ....

[Article contains additional citation context not shown here]

R. Sarpeshkar, W. Bair and C. Koch, "Visual Motion Computation in Analog VLSI using Pulses," in Advances in Neural Information Processing Systems 5, D. S. Touretzky, Ed. San Mateo, CA: Morgan Kaufmann, 1993, pp. 781-788.

....tuned to different velocities, or a single array with adaptive tuning, would be necessary to measure velocity over an extended range. This scheme is therefore not suitable for the monolithic implementation of dense velocity sensing arrays. A binary correlation scheme was implemented in a 1 D array [31]. It used a spatial edge detector, that convolved the image with a difference of exponentials kernel, implemented with two resistive grids with different resistances [32] An edge of sufficiently high contrast triggered a voltage pulse of fixed amplitude and width. This pulse was then timed ....

....in dense arrays. Two time of travel algorithms were implemented with more compact circuits, that unambiguously encoded 1 D velocity over considerable velocity, contrast, and illumination ranges. Both used a temporal edge detector as input stage, that responded to ON edges [9, 10] In one circuit [10, 31], an edge signal generated a voltage pulse of fixed amplitude and width at each pixel location. The pulses from two adjacent locations were fed into two motion circuits, one for each direction. For motion in the preferred direction, such a motion circuit output a pulse whose width corresponded to ....

R. Sarpeshkar, W. Bair and C. Koch, "Visual Motion Computation in Analog VLSI using Pulses," in Advances in Neural Information Processing Systems 5, D. S. Touretzky, Ed. San Mateo, CA: Morgan Kaufman, pp. 781-788, 1993.

No context found.

R. Sarpeshkar, W. Bair, and C. Koch. Visual motion computation in analog VLSI using pulses. Neural information processing systems 5, S. Hanson, J. Cowan, C. Giles editors. Morgan Kaufman, San Mateo, pp. 781-788, 1993.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC