{"title": "A Low-Power Analog VLSI Visual Collision Detector", "book": "Advances in Neural Information Processing Systems", "page_first": 987, "page_last": 994, "abstract": "", "full_text": " \n\n \n \n \n \n \n \n\nA Low-Power Analog VLSI Visual \n\nCollision Detector \n\nDepartment of Electrical and Computer Engineering \n\n \n\n \n\n \n \n   \n\nReid R. Harrison \n\nUniversity of Utah \n\nSalt Lake City, UT 84112 \nharrison@ece.utah.edu \n\nAbstract \n\nWe have designed and tested a single-chip analog VLSI sensor that \ndetects  imminent  collisions  by  measuring  radially  expansive  optic \nflow.    The  design  of  the  chip  is  based  on  a  model  proposed  to \nexplain  leg-extension behavior in flies during  landing  approaches.  \nA new elementary motion detector (EMD) circuit was developed to \nmeasure optic flow.  This EMD circuit models the bandpass nature \nof  large  monopolar  cells  (LMCs)  immediately  postsynaptic  to \nphotoreceptors  in  the  fly  visual  system.    A  16  \u00d7  16  array  of  2-D \nmotion detectors was fabricated on a 2.24 mm \u00d7 2.24 mm die in a \nstandard  0.5-\u00b5m  CMOS  process.    The  chip  consumes  140  \u00b5W  of \npower from a 5 V supply.  With the addition of wide-angle optics, \nthe sensor is able to detect collisions around 500 ms before impact \nin complex, real-world scenes. \n\n1  Introduction \n\nMany animals (cid:150) from flies to humans (cid:150) are capable of visually detecting imminent \ncollisions  caused  either  by  a  rapidly  approaching  object  or  self-motion  towards  an \nobstacle.  Neurons dedicated to this task have been found in the locust [1] and the \npigeon  [2].    Borst  and  Bahde  have  shown  that  flies  use  visual  information  to  time \nthe extension of their legs on landing approaches [3].   \nWhile several models have been proposed to explain collision detection, the model \nproposed  in  [3]  is  particularly  amenable  to  hardware  implementation.    The  model, \nshown  in  Fig.  1,  employs  a  radially-oriented  array  of  motion  detectors  centered  in \nthe direction of flight.  As the animal approaches a static object, an expansive optic \nflow field is produced on the retina.  A wide angle field of view is useful since optic \nflow  in  the  direction  of  flight  will  be  zero.    The  response  of  this  radial  array  of \nmotion detectors is summed and then passed through a leaky integrator (a lowpass \nfilter).  If this response exceeds a fixed threshold, an imminent collision is detected \nand  the  animal  can  take  evasive  action  or  prepare  for  a  landing.    This  expansive \noptic flow model has recently been used to explain landing and collision avoidance \nresponses  in  the  fruit  fly  [4].    A  similar  algorithm  has  been  implemented  in  a \ntraditional  CPU  for  autonomous  robot  navigation  [5].    In  this  work,  we  present  a \nsingle-chip analog VLSI sensor developed to implement this model. \n\n \n\n\f \n\nField of View \n\nradially-oriented \n\nelementary \n\nD \n\nmotion detectors \n\n(EMDs) \n\nspatial \n\nsummation \n\nleaky \n\nintegrator \n\n\u03c4 = RC \n\nthreshold \n\ncomparator \n\ncollision \ndetect \n\n \n\n \n\nFigure 1: Diagram of collision detection algorithm. \n\n2  Elementary Motion Detectors \n\nOur  collision  detection  algorithm  uses  an  array  of  radially-oriented  elementary \nmotion detectors (EMDs) to sense image expansion.  Simulations by the author have \nshown that the structure and properties of the EMDs strongly affect the accuracy of \nthis algorithm [6].  We use an enhanced version of the familiar delay-and-correlate \nor  (cid:147)Reichardt(cid:148)  EMD  first  proposed  by  Hassenstein  and  Reichardt  in  the  1950s  to \nexplain the optomotor response of beetles [7].  Fig. 2 shows a diagram of the EMD \nused in our collision sensor. \nThe first stage of the EMD is photoreception, where light intensity is transduced to \na signal vphoto.  Since light intensity is a strictly positive value, the mean intensity of \nthe  scene  must  be  subtracted.    Since  we  are  interested  in  motion,  it  is  also \nadvantageous to amplify transient signals. \nSuppressing dc illumination and enhancing ac components of photoreceptor signals \nis  a  common  theme  in  many  biological  visual  systems.    In  flies,  large  monopolar \ncells  (LMCs)  directly  postsynaptic  to  photoreceptors  exhibit  transient  biphasic \nimpulse responses approximately 40-200  ms in duration [8], [9].  In the  frequency \ndomain, this can be seen as a bandpass filtering operation that attenuates dc signals \nwhile  amplifying  signals  in  the  2-40  Hz  range  [9],  [10].    In  the  lateral  geniculate \nnucleus of cats, (cid:147)lagged(cid:148) and (cid:147)non-lagged(cid:148) cells exhibit transient biphasic impulse \nresponses 200-300 ms in duration and act as bandpass filters amplifying signals in \nthe  1-10  Hz  range  [11].    This  filtering  has  recently  been  explained  in  terms  of \ntemporal decorrelation, and can be seen as way of removing redundant information \nfrom the photoreceptor signal before further processing [9], [12]. \nAfter  this  (cid:147)transient  enhancement(cid:148),  or  temporal  decorrelation,  the  signals  are \ndelayed  using  the  phase  lag  of  a  lowpass  filter.    While  not  a  true  time  delay,  the \nlowpass filter matches data from animal experiments and makes the Reichardt EMD \nequivalent  to  the  oriented  spatiotemporal  energy  filter  proposed  by  Adelson  and \nBergen [13].  Before correlating the adjacent delayed and  non-delayed  signals,  we \napply a saturating static nonlinearity to each channel.  Without such a nonlinearity, \nthe delay-and-correlate EMD exhibits a quadratic dependence on image contrast.  In \nfly tangential neurons, motion responses show a quadratic dependence only at very \nlow  contrasts,  then  quickly  become  largely  independent  of  image  contrast  for \ncontrasts above 30%.  Egelhaaf and Borst proposed the presence of this nonlinearity \nin the biological EMD to explain this contrast independence [14].  Functionally, it is \nnecessary to prevent high-contrast edges from dominating the summed output of the \nEMD array. \n\n \n\n\f \n\n \n\nimage motion \n\nFig. 3 \n\nvphoto-L \n\nvphoto-R \n\nLMC \n\nLMC \n\nphotoreceptors \n\ntemporal decorrelation \n(Large Monopolar Cells) \n\nFig. 4 \n\nvLMC-L \n\ndelay \n\nvdelay-L \n\nvLMC-R \n\ndelay \n\ndelay \n\nvdelay-R \n\niout-L \n\niout-R \n\niout \n\nsaturating nonlinearity \n\ncorrelation (multiplication) \n\nopponent subtraction \n\n \n\nFigure 2: Elaborated delay-and-correlate elementary motion detector (EMD) \n\nAfter  correlation,  opponent  subtraction  produces  a  strong  directionally  selective \nsignal that is taken as the output of the EMD.  Unlike algorithms that find and track \nfeatures  in  an  image,  the  delay-and-correlate  EMD  does  not  measure  true  image \nvelocity  independent  of  the  spatial  structure  of  the  image.    However,  recent  work \nhas shown that for natural scenes, these Reichardt EMDs give reliable estimates of \nimage velocity [15].  This reliability is improved by the addition of LMC bandpass \nfilters  and  saturating  nonlinearities.    Experiments  using  earlier  versions  of  silicon \nEMDs  have  demonstrated  the  ability  of  delay-and-correlate  motion  detectors  to \nwork at very low signal-to-noise ratios [16]. \n\n3  Integrated Circuit Implementation \n\nWe  adapted  the  EMD  shown  in  Fig.  2  to  a  small,  low-power  CMOS  integrated \ncircuit.  Fig. 3 shows a schematic of the photoreceptor and LMC bandpass filter.  A \n35  (cid:181)m  (cid:215)  35  (cid:181)m  well-substrate  photodiode  with  diode-connected  pMOS  load \nconverts the diode photocurrent into a voltage vphoto that is a logarithmic function of \nlight intensity.  A pMOS source follower biased by ISF = 700 pA buffers this signal \nso that the input capacitance of the LMC circuit does not load the photoreceptor. \nThe  LMC  bandpass  filter  consists  of  two  operational  transconductance  amplifiers \n(OTAs) and three capacitors.  The OTAs in the circuit are implemented with pMOS \ndifferential  pairs  using  diode-connected  transistors  for  source  degeneration  for \nextended linear range (see inset, Fig. 3).  The transfer function of the LMC circuit is \ngiven by \n\nv\n\n( )\ns\nLMC\n( )\nv\ns\n\nin\n\n \n\nwhere \n\n\u22c5\u2212=\n\n\u03c4\nsNA\n0\n)\n2\ns\n\n(\n\u03c4\n1\n\n(\n)\n\u2212\u22c5\n\u03c4\ns\n1\n0\n\u03c4\ns\n+\n+\n1\n1\nQ\n\n\u2212=\n\nAN\n\u03b2\n\n\u22c5\n\n\u03c4\n1\n\n\uf8eb\n\uf8ec\uf8ec\n\uf8ed\ns\n\n\u2212\n\n1\n\n+\n\n\u03c4\n1\n\u03b2\n1\nQ\n\n+\n\ns\n\n\uf8f6\n\uf8f7\uf8f7\n\uf8f8\n1\n\u03c4\ns\n1\n\n \n\nC=0\u03c4\nmg\n)\n\u2248\n\u2212+\n1\n\nN\n\n(\nAN\u03b2\n\n=\n\n)(\nK\n1\n\n+\n\nNAK\n\n if \n\nKA\n,\n\n>>\n\n1\n\n \n\n \n\n \n\n \n\n \n\n(1) \n\n(2) \n\n(3) \n\n\fC \n\ngm /N \n\ngm \n\nISF \n\nvin \n\nAC \n\nVREF \n\nvphoto \n\n \n\nv- \n\nv+ \n\ngm \n\nout \n\nIB \n\ngm = \u03baIB/2UT \n\nv+ \n\nv- \nout \n\nvLMC \n\nKC \n\n \n\nFigure 3: Schematic of photoreceptor/LMC circuit.  Detail of operational \n\ntransconductance amplifier (OTA) shown in inset. \n\n \n\n1 \u03b2\u03c4\n\u03c4 =\n0\n\u03b2\n(\n)NK\n+\n\n=\n\nQ\n\n \n\n(4) \n\n(5) \n\n \n\n \n\n \n\nThe output signal vLMC is centered around VREF, a dc voltage which was set to 1.0 V.  \nWe sized the capacitors in  our circuit to give A = 20 and K = 5 (with  C = 70 fF).  \nThe transconductance of the lower OTA was set by adjusting its bias current IB: \n\n \n\n=\n\ng\n\nm\n\n\u03ba\n(\n\u03ba\n+\n\n\u22c5\n\nI\nB\n)\nU\n21\n\nT\n\n \n\n(6) \n\nwhere  \u03ba  is  the  weak  inversion  slope  (typically  between  0.6  and  0.9)  and  UT  is  the \nthermal voltage kT/q (approximately 26 mV at room temperature).  We set the bias \ncurrent in the upper OTA five times smaller to achieve N = 5. \nAs we see from (1), the LMC circuit acts as an ac-coupled bandpass filter centered \nat f1 = 1/2\u03c0\u03c41, with a quality factor Q set to 2.5 by capacitor and current ratios.  The \ncircuit  also  has  a  zero  at  \u03b2f1,  but  since  \u03b2  =  25  in  our  circuit,  the  zero  takes  effect \noutside that passband and thus has little practical effect on the filter.  We used a bias \ncurrent of IB = 35 pA in the lower  OTA and 7 pA  in  the  upper OTA to center the \npassband  near  20  Hz,  which  was  chosen  because  it  lies  in  the  range  of  LMC \nresponse  measured  in  the  fly. \n  This  LMC  circuit  represents  a  significant \nimprovement  over  a  previous  silicon  EMD  design,  which  used  only  a  first-order \nhighpass  filter  to  block  dc  illumination  [16].    The  LMC  circuit  presented  here \nallows  the  designer  to  adjust  the  center  frequency  and  Q  factor  to  selectively \namplify frequencies present in moving images. \nThe LMC circuits  from each photoreceptor pass  their signals to the  the delay-and-\ncorrelate circuit shown in Fig. 4.  The delay is implemented as a first-order lowpass \nfilter.    The  OTAs  in  this  circuit  used  two  diode-connected  transistors  in  series  for \nextended linear range.  The time constant of this filter is given by \n\n\u03c4\nLPF\n\n=\n\nC\ng\n\nm\n\nLPF\n\u2212\n\nLPF\n\n \n\n(7) \n\n \n\n \n\n\f \n\n \n\nvdelay-L \n\ngm-LPF \n\nCLPF \n\nImult \n\nvLMC-L  vLMC-R \n\nVREF \n\nVREF \n\ngm-LPF \n\nvdelay-R \n\nCLPF \n\nImult \n\nVW \n\nVREF \n\nVW \n\nVW \n\nVREF \n\nVW \n\niout-L- \n\niout-R- \n\niout-L+ \n\niout+ \n\niout-R+ \n\niout- \n\n \n\nFigure 4: Schematic of delay-and-correlate circuit.  OTA-based gm-C filters are used \nas low-pass filters.  Subthreshold CMOS Gilbert multipliers are used for correlation. \nWe  used  CLPF  =  700  fF  and  set  \u03c4LPF  to  around  25  ms,  which  is  in  the  range  of \nbiological motion detectors.  This required a bias current of 9 pA for each OTA. \nWe  implemented  the  correlation  function  using  a  CMOS  Gilbert  multiplier \noperating in subthreshold [17].  The output currents of the multipliers in Fig. 4 can \nbe expressed as: \n\n(\n\u03ba\nv\n\n \n\n \n\ni\n\noutL\n\n+\n\n\u2212\n\ni\n\noutL\n\n\u2212\n\n=\n\nI\n\nmult\n\ntanh\n\ni\n\noutR\n\n+\n\n\u2212\n\ni\n\noutR\n\n\u2212\n\n=\n\nI\n\nmult\n\ntanh\n\n(\n\u03ba\nv\n\n\u2212\ndelay\nL\nU\n2\n\n\u2212\nR\ndelay\nU\n2\n\n)\n\n\u2212\n\nV\n\nREF\n\n(\n\u03ba\nv\n\ntanh\n\n)\n\nV\n\nREF\n\n(\n\u03ba\nv\n\ntanh\n\nT\n\u2212\n\nT\n\n\u2212\n\nV\n\nREF\n\n)\n\n)\n\nV\n\nREF\n\nT\n\u2212\n\nT\n\n\u2212\nLMC\nR\nU\n2\n\n\u2212\nL\nLMC\nU\n2\n\n \n\n \n\n(8) \n\n(9) \n\nFor  small  differential  input  voltages,  tanh(x)  \u2248  x  and  the  circuit  acts  as  a  linear \nmultiplier.  As the input signals grow larger, the tanh nonlinearity dominates and the \ncircuit  acts  more  like  a  digital  exclusive-or  gate.    We  use  this  inherent  circuit \nnonlinearity  as  the  desired  saturating  nonlinearity  in  our  EMD  model  (see  Fig.  1).  \nThe  previous  LMC  circuit  provides  sufficient  gain  to  ensure  that  we  are  usually \noperating well outside the linear range of the multipliers. \nTraditional  CMOS  Gilbert  multipliers  require  that  the  dc  level  of  the  upper \ndifferential  input  be  shifted  relative  to  the  dc  level  of  the  lower  differential  input.  \nThis is required to keep the transistors in saturation.  To avoid the cost in chip area, \npower  consumption,  and  mismatch  associated  with  level  shifters,  we  introduce  a \nnovel circuit modification that allows both the upper and lower differential inputs to \noperate  at  the  same  dc  level.    We  lower  the  well  potential  of  the  lower  pMOS \ntransistors  from  VDD  to  a  dc  voltage  VW  (see  Fig.  4).    This  lowered  well  voltage \ncauses the sources of these  transistors to operate at a lower potential,  which keeps \nthe upper transistors in saturation.  We use VW = 2.5 V in our circuit.  (Care must be \ntaken not to make VW too low, as parasitic source-well-substrate pnp transistors can \nbe activated.) \n\n \n\n\f \n\n \n\n52\u00b0\u00b0\u00b0\u00b0 \n\n74\u00b0\u00b0\u00b0\u00b0 \n\n \n\nFigure 5: EMD pattern on chip.  Ultra-wide-angle optics gave the chip a field of \n\nview ranging from \u201352\u00b0 to \u201374\u00b0. \n\nThe output of the Gilbert  multiplier is a differential current.  The signals from  the \nleft  and  right  correlators  are  easily  subtracted  by  summing \ntheir  currents \nappropriately.  Similarly, current summation on two global wires is used to sum the \nmotion signals over the entire EMD array. \n\n4  Experimental Results \n\nWe  fabricated  a  16  (cid:215)  16  EMD  array  in  a  0.5-(cid:181)m  2-poly,  3-metal  standard  CMOS \nprocess.  The 2.24 mm (cid:215) 2.24 mm die contained a 17 (cid:215) 17 array of (cid:147)pixels,(cid:148) each \nmeasuring 100 (cid:181)m (cid:215) 100 (cid:181)m.  Each pixel contained a photoreceptor, LMC circuit, \nlowpass  (cid:147)delay(cid:148)  filter,  and  four  correlators.    These  correlators  were  used  to \nimplement two independent EMDs: a vertical motion detector connected to the pixel \nbelow  and  a  horizontal  motion  detector  connected  to  the  pixel  to  the  right.    The \noutput signals from a subset of the EMDs representing radial outward motion were \nconnected to two global wires, giving a differential current signal that was taken off \nchip on two pins. \nFig.  5  shows  the  EMDs  that  were  summed  to  produce  the  global  radial  motion \nsignal.  Diagonally-oriented EMDs were derived from the sum of a horizontal and a \nvertical EMD.  The center 4 (cid:215) 4 pixels  were ignored, as motion near the center of \nthe  field  of  view  is  typically  very  small  in  collision  situations.    We  used  custom-\nbuilt ultra-wide-angle optics to  give the chip a  field of  view  ranging  from \u201352\u00b0 at \nthe sides to \u201374\u00b0 at the corners.  Simulations revealed that a field of view of around \n\u201360\u00b0 was necessary for reasonable performance using this algorithm [6]. \nBefore testing the array,  we characterized an individual LMC circuit configured to \nhave  a  voltage  input  vphoto  provided  from  off  chip  using  a  function  generator.    We \nprovided  a  1.4  Hz,  100  mVpp  square  wave  and  observed  the  LMC  circuit  output.  \nAs shown in Fig. 6a, the LMC circuit exhibits a transient oscillatory step response \nsimilar  to  its  biological  counterpart.    Using  a  spectrum  analyzer,  we  measured  the \ntransfer  function  of  the  circuit  (see  Fig.  6b).    The  LMC  circuit  acts  as  a  bandpass \nfilter centered at 19 Hz, with a measured Q of 2.3. \n\n \n\n\f \n\n \n\nFigure 6: Measurement of LMC circuit performance.  (a) Step response of LMC \n\ncircuit.  (b) Frequency tuning of LMC circuit. \n\nThe  entire  chip  consumed  140  \u00b5W  of  power.    Most  of  this  was  consumed  by \nperipheral  biasing  circuits;  the  17 \u00d7  17  pixel  array  used  only  5.2  \u00b5W  (18  nW  per \npixel).    To  test  the  complete  collision  detection  chip,  we  implemented  the  leaky \nintegrator (\u03c4leak = 50 ms) and comparator from Fig. 1 using off-chip components.  In \nfuture implementations, these circuits could be built on chip using little power. \nWe tested the chip by mounting it on a small motorized vehicle facing forward with \nthe lens centered 11 cm above the floor.  The vehicle traveled in a straight path at 28 \ncm/s.  Fig. 7 shows the output from the leaky integrator as the chip moves across the \nfloor and collides with the center of a 38 cm \u00d7 38 cm trash can in our lab.  The peak \nresponse  of \nthe  chip  occurs  approximately  500  ms  before  contact,  which \ncorresponds to a distance of 14 cm.  At this point, the edges of the trash can subtend \nan angle of 54\u00b0.  After this point, the edges of the can move beyond the chip(cid:146)s field \nof view, and the response decays rapidly.  The rebound in response observed in the \nlast  100  ms  may  be  due  to  the  chip  seeing  the  expanding  shadow  cast  by  its  own \nlens on the side of the can just before contact.  \n\n5  Conclusions \n\nThe response of our chip, which peaks and then collapses before impact, is similar \nto  activity  patterns  observed  in  the  LGMD  neuron  in  locusts  [1]  and  \u03b7  neurons  in \npigeons [2] during  simulated collisions.  While  more  complex  models positing the \nmeasurement of true image velocity and object size have been used to explain this \npeculiar time course [1], we observe that a simple model integrating the output of a \nradial EMD array gives qualitatively similar responses. \nWe have demonstrated that this model of collision detection can be implemented in \na  small,  low-power,  single-chip  sensor.    Further  testing  of  the  chip  on  mobile \nplatforms should better characterize its performance. \n\nAc knowledg me nts \nThis work was partially supported by a contract from the Naval Air Warfare Center, \nChina Lake, CA. \n\nReferences \n[1] F. Gabbiani, H.G.  Krapp, and G.  Laurent, (cid:147)Computation of object approach by  a  wide-\nfield, motion-sensitive neuron,(cid:148) J. Neurosci. 19:1122-1141, 1999. \n\n \n\n\f \n\n \n\nFigure 7: Measured output of collision detection chip. \n\n[2] H. Sun and B.J. Frost, (cid:147)Computation of different optical variables of looming objects in \npigeon nucleus rotundus neurons,(cid:148) Nature Neurosci. 1:296-303, 1998. \n[3]  A.  Borst  and  S.  Bahde,  (cid:147)Visual  information  processing  in  the  fly(cid:146)s  landing  system,(cid:148)  J. \nComp. Physiol. A 163:167-173, 1988. \n [4]  L.T.  Tammero  and  M.H.  Dickinson,  (cid:147)Collision-avoidance  and  landing  responses  are \nmediated by separate pathways in the fruit fly, Drosophila melanogaster,(cid:148) J. Exp. Biol.205: \n2785-2798, 2002. \n[5]  A.P.  Duchon,  W.H.  Warren,  and  L.P.  Kaelbling,  (cid:147)Ecological  robotics,(cid:148)  Adaptive \nBehavior 6:473-507, 1998. \n[6]  R.R.  Harrison,  (cid:147)An  algorithm  for  visual  collision  detection  in  real-world  scenes,(cid:148) \nsubmitted to NIPS 2003. \n[7] B. Hassenstein and W. Reichardt, (cid:147)Systemtheoretische Analyse der Zeit-, Reihenfolgen-, \nund  Vorzeichenauswertung  bei  der  Bewegungsperzeption  des  R(cid:252)sselk(cid:228)fers  Chlorophanus,(cid:148) \nZ. Naturforch. 11b:513-524, 1956. \n[8]  S.B.  Laughlin,  (cid:147)Matching  coding,  circuits,  cells,  and  molecules  to  signals  (cid:150)  general \nprinciples  of  retinal  design  in  the  fly(cid:146)s  eye,(cid:148)  Progress  in  Ret.  Eye  Research  13:165-196, \n1994. \n[9]  J.H.  van  Hateren,  (cid:147)Theoretical  predictions  of  spatiotemporal  receptive  fields  of  fly \nLMCs, and experimental validation,(cid:148) J. Comp. Physiol. A 171:157-170, 1992. \n[10] J.H. van Hateren, (cid:147)Processing of natural time series of intensities by the visual system \nof the blowfly,(cid:148) Vision Res. 37:3407-3416, 1997. \n[11] A.B. Saul and A.L. Humphrey, (cid:147)Spatial and temporal response properties of lagged and \nnonlagged cells in cat lateral geniculate nucleus,(cid:148) J. Neurophysiol. 64:206-224, 1990. \n[12] D.W. Dong and J.J. Atick, (cid:147)Temporal decorrelation: a theory of lagged and nonlagged \nresponses in the lateral geniculate nucleus,(cid:148) Network 6:159-178, 1995. \n[13]  E.H.  Adelson  and  J.R.  Bergen,  (cid:147)Spatiotemporal  energy  models  for  the  perception  of \nmotion,(cid:148) J. Opt. Soc. Am. A 2:284-299, 1985. \n[14] M. Egelhaaf and A. Borst, (cid:147)Transient and steady-state response properties of movement \ndetectors,(cid:148) J. Opt. Soc. Am. A 6:116-127, 1989. \n[15]  R.O.  Dror,  D.C.  O(cid:146)Carroll,  and  S.B.  Laughlin,  (cid:147)Accuracy  of  velocity  estimation  by \nReichardt correlators,(cid:148) J. Opt. Soc. Am. A 18:241-252, 2001. \n[16] R.R. Harrison and C. Koch, (cid:147)A robust analog VLSI Reichardt motion sensor,(cid:148) Analog \nIntegrated Circuits and Signal Processing 24:213-229, 2000. \n[17] C. Mead, Analog VLSI and Neural Systems, Reading, MA: Addison-Wesley, 1989. \n\n \n\n\f", "award": [], "sourceid": 2492, "authors": [{"given_name": "Reid", "family_name": "Harrison", "institution": null}]}