{"title": "Visual gesture-based robot guidance with a modular neural system", "book": "Advances in Neural Information Processing Systems", "page_first": 903, "page_last": 909, "abstract": null, "full_text": "Visual gesture-based robot  guidance \n\nwith a  modular  neural system \n\nE.  Littmann, \n\nAbt.  Neuroinformatik,  Fak.  f.  Informatik \n\nUniversitat  Ulm,  D-89069  Ulm,  FRG \n\nenno@neuro.informatik.uni-ulm.de \n\nA.  Drees, and H.  Ritter \n\nAG  Neuroinformatik,  Techn.  Fakultat \nUniv.  Bielefeld,  D-33615  Bielefeld,  FRG \n\nandrea,helge@techfak.uni-bielefeld.de \n\nAbstract \n\nWe report on the development of the modular neural system  \"SEE(cid:173)\nEAGLE\"  for  the  visual  guidance  of robot  pick-and-place  actions. \nSeveral  neural  networks  are  integrated  to  a  single system  that vi(cid:173)\nsually  recognizes  human hand  pointing gestures  from  stereo  pairs \nof color  video  images.  The output of the  hand recognition stage is \nprocessed  by  a  set  of color-sensitive  neural  networks  to  determine \nthe cartesian location of the target object that is  referenced  by the \npointing gesture.  Finally, this information is  used  to guide a robot \nto grab  the  target  object  and  put  it  at  another  location  that  can \nbe specified  by a second pointing gesture.  The accuracy of the cur(cid:173)\nrent  system allows to identify the location of the referenced  target \nobject to an accuracy of 1 cm  in  a  workspace  area of 50x50 cm.  In \nour  current  environment,  this  is  sufficient  to  pick  and  place  arbi(cid:173)\ntrarily positioned target objects within the workspace.  The system \nconsists  of neural  networks  that  perform  the  tasks  of image seg(cid:173)\nmentation, estimation of hand  location, estimation of 3D-pointing \ndirection, object recognition, and necessary  coordinate transforms. \nDrawing heavily on the use of learning algorithms, the functions of \nall  network modules were  created from  data examples only. \n\nIntroduction \n\n1 \nThe  rapidly  developing  technology  in  the  fields  of robotics  and  virtual  reality  re(cid:173)\nquires  the  development  of new  and  more  powerful  interfaces  for  configuration  and \ncontrol  of such  devices.  These  interfaces should  be intuitive for  the human advisor \nand  comfortable  to  use.  Practical  solutions  so  far  require  the  human  to  wear  a \ndevice that can transfer the necessary information.  One typical example is the data \nglove  [14,  12].  Clearly,  in  the  long  run solutions  that are  contactless  will  be  much \nmore  desirable,  and  vision  is  one  of the  major  modalities  that  appears  especially \nsuited for  the  realization of such solutions. \nIn the present  paper, we focus  on a still restricted  but very important task in robot \ncontrol, the guidance of  robot pick-and-place actions by unconstrained human poin(cid:173)\nting gestures in a realistic laboratory environment.  The input of target locations by \n\n\f904 \n\nE.  LITTMANN, A. DREES, H.  RITTER \n\npointing gestures  provides  a  powerful,  very  intuitive and comfortable functionality \nfor  a  vision-based  man-machine interface  for  guiding  robots  and  extends  previous \nwork that focused on the detection of hand location or the discrimination of a small, \ndiscrete  number of hand gestures only  [10,  1,  2,  8].  Besides  two  color  cameras,  no \nspecial  device  is  necessary  to evaluate the gesture  of the human operator. \nA  second  goal  of our  approach  is  to  investigate  how  to  build  a  neural  system  for \nsuch  a  complex  task  from  several  neural  modules.  The  development  of advanced \nartificial  neural systems  challenges  us with the task of finding  architect.ures  for  the \ncooperat.ion  of multiple functional  modules  such  that.  part  of the  structure  of the \noverall system can  be  designed  at a useful level of abstraction, but at the same t.ime \nlearning  can  be  used  to create  or fine-tune  the functionality of parts of t.he  system \non  the basis of suit.able training examples. \nTo  approach  this  goal  requires  to shift  the  focus  from  exploring  t.he  properties  of \nsingle  networks  to  exploring  the  propert.ies  of entire  systems  of neural  networks. \nThe work on  \"mixtures of experts\"  [3,  4]  is  one important contribution along  these \nlines.  While  this  is  a  widely  applicable  and  powerful  approach,  there  clearly  is \na  need  to  go  beyond  the  exploration  of strictly  hierarchical  systems  and  to  gain \nexperience  with  architectures  t.hat  admit more  complex types  of information flow \nas  required  e.g.  by  the  inclusion  of feat.ures  such  as  control  of focal  attention  or \nreent.rant  processing  branches.  The  need  for  such  features  arose  very  naturally in \nthe context of the task  described  above,  and  in the following sect.ion  we  will  report \nour results  wit.h  a system architecture  that is  crucially based on the exploitation of \nsuch  elements. \n2  System architecture \nOur system, described  in fig.  1,  is situated in  a complex laboratory environment.  A \nrobot  arm with  manipulator is  mounted at one side  of a  table with  several  objects \nof different  color  placed on it.  A  human operator  is  positioned at  the  next side to \nthe right of the robot.  This scenery  is  watched  by  two  cameras from the other two \nsides from high above.  The cameras yield  a stereo  color image of t.he  scene  (images \n10).  The operator points with one  hand  at one of the objects on the table.  On the \nbasis of the  image information, the object  is located  and the  robot grabs  it.  Then, \nthe operator  points at another  location, where  the robot releases  the object. 1 \nThe syst.em consists of several  hardware components:  a  PUMA  560 robot  arm with \nsix  axes  and a  three-fingered  manipulator 2;  two single-chip  PULNIX  color  cameras; \ntwo  ANDRox  vision  boards  with  software  for  data  acquisition  and  processing;  a \nwork  space  consisting  of a  table with  a  black grid  on  a  yellow  surface.  Robot  and \nperson  refer  to  the  same  work  space.  Bot.h  cameras  must  show  both  the  human \nhand  and  the  table  with  the  objects.  Within  this  constraint,  the  position  of the \ncameras can  be  chosen  freely  as  long as they yield  significantly different  views. \nAn  important  prerequisite  for  the  recognition  of the  pointing  direction  is  the  seg(cid:173)\nmentation of the human hand from  the background scenery.  This task is  solved  by \na  LLM  network  (Sl)  trained  to  yield  a  probability value  for  each  image  pixel  to \nbelong  to  the  hand  region.  The  training  is  based  on  t.he  local  color  information. \nThis procedure  has been  investigated in  [7]. \nAn  important feature  of the  chosen  method is  the great  reliability and  robustness \nof both the classification performance and the localization accuracy of the searched \nobject.  Furthermore, the performance is  quite constant over  a wide range of image \nresolutions.  This allows a fast two-step  procedure:  First, the images are segmented \nin  low  resolution  (Sl:  11  -+ A1)  and  the  hand position is  extracted.  Then,  a small \n\n1 In analogy to the sea eagle who watches its prey from  high  above,  shoots down to grab \n\nthe  prey,  and  then flies  to  a safe  place  to feed,  we  nicknamed  our system  \"SEE-EAGLE\". \n\n2Development  by  Prof.  Pfeiffer,  TV  Munich \n\n\fVisual Gesture-based Robot Guidance with a Modular Neural System \n\n905 \n\nFig.  1:  System architecture.  From  two color camera images 10 we extract the hand position \n(11  I>  Sl  I>  A1  (pixel  coord.)  I>  P1  I>  cartesian  hand  coord.).  In  a  subframe  centered  on \nthe hand location  (12)  we  determine  the pointing  direction  (12  I>  S2 I>  A2 (pixel  coord.)  I> \nG  I>  D  I>  pointing  angles).  Pointing  direction  and  hand location  define  a  cartesian  target \nlocation  that is  mapped  to image coord.  that define  the centers of object subframes (10  I> \nP2 I>  13).  There we  determine  the  target  object  (13  I>  S3  I>  A3) and  map  the pixel coord. \nof its centers  to  world  coord.  (A3  I>  P3 I>  world  target loc.).  These coordinates  are  used \nto guide  the robot  R  to  the target object. \n\n\f906 \n\nE.  LITTMANN. A. DREES. H.  RlTIER \n\nsubframe  (12)  around  the  estimated  hand  position  is  processed  in  high  resolution \nby another dedicated  LLM  network  (S2:  12  - t  A2).  For details of the segmentation \nprocess,  refer  to [6]. \nThe  extraction  of  hand  information  by  LLMs  on  the  basis  of  Gabor  masks  has \nalready been studied for  hand posture  [9]  and orientation [5].  The method is  based \non  a segmented image containing the  hand only  (A2).  This  image is  filtered  by  36 \nGabor  masks  that  are  arranged  on  a  3x3  grid  with  4  directions  per  grid  position \nand centered on the hand.  The filter  kernels have a radius of 10  pixels,  the distance \nbetween  the  grid  points  is  20  pixels.  The  36  filter  responses  (G)  form  the  input \nvector for  a LLM  network  (D).  Further details of the processing  are reported  in  [6]. \nThe  network  yields  the  pointing  direction  of the  hand  (D:  12  - t  G  - t  pointing \ndirection).  Together  with  the  hand  position  which  is  computed  by  a  parametrized \nself-organizing  map  (\"PSOM\",  see  below  and  [11,  13])  (P1:  Al  - t  cartesian  hand \nposition),  a  (cartesian)  target  location  in  the  workspace  can  be  calculated.  This \nlocation  can  be  retransformed  by  the  PSOM  into  pixel  coordinates  (P2:  cartesian \ntarget location  - t  target pixel  coordinates).  These  coordinates  define  the center  of \nan  \"attention  region\"  (13)  that  is  searched  for  a  set  of predefined  target  objects. \nThis object  recognition is  performed  by  a  set  of LLM  color segmentation networks \n(S3:  13  - t  A3),  each  previously  trained  for  one  of the  defined  targets.  A  ranking \nprocedure is used to determine the target object.  The pixel coordinates ofthe target \nin the segmented image are mapped by  the PSOM  to world  coordinates (P3:  A3  - t  \ncartesian target position).  The robot R now moves to above these world coordinates, \nmoves vertically down,  grabs whatever is there,  and moves upward again.  Now,  the \nsystem evaluates  a  second  pointing gesture  that specifies  the  place  where  to place \nthe object.  This time, the world coordinates calculated on the basis of the pointing \ndirection  from  network  D  and  the  cartesian  hand  location  from  PSOM  PI  serve \ndirectly  as  target location for  the  robot. \nFor  our  processing  we  must  map  corresponding  pixels  in the stereo  images to  car(cid:173)\ntesian  world  coordinates.  For  these  transformations,  training  data was  generated \nwith  aid  of the  robot  on  a  precise  sampling  grid.  We  automatically extract  the \npixel  coordinates of a  LED  at the  tip  of the robot  manipulator from  both images. \nThe  seven-dimensional  feature  vector  serves  as  training  input  for  an  PSOM  net(cid:173)\nwork [11].  By  virtue of its capability to represent  a transformation in  a symmetric, \n\"multiway\" -fashion, this offers the additional benefit that both the camera-to-world \nmapping and its inverse  can be obtained with a single network trained only once on \na  data set  of 27  calibration  positions  of the  robot.  A  detailed  description  for  such \na  procedure  can  be  found  in  [13]. \n\n3  Results \n3.1  System performance \nThe  accuracy  of the  current  system  allows  to  estimate  the  pointing  target  to  an \naccuracy  of 1 \u00b1  0.4 cm  (average  over  N  =  7 objects  at  randomly chosen  locations \nin  the  workspace)  in  a  workspace  area of 50x50 cm.  In  our  current  environment, \nthis  is  sufficient  to  pick  and  place  any  of the  seven  defined  target  objects  at  any \nlocation in  the workspace.  This accuracy  can  only be  achieved  if we  use  the object \nrecognition module described  in sec.  2.  The output of the pointing direction module \napproximates the target location with an considerably lower accuracy of 3.6\u00b1 1.6 cm. \n\nImage segmentation \n\n3.2 \nThe problem to evaluate these preprocessing steps has been discussed previously [7], \nespecially the relation of specifity  and sensitivity of the  network for  the given  task. \nAs  the pointing recognition  is  based on  a subframe centered on the hand center,  it \nis very  sensitive to deviations from this center  so that a good localization accuracy \n\n\fVisual Gesture-based Robot Guidance with a Modular Neural System \n\n907 \n\nis  even  more  important than  the  classification  rate.  The  localization  accuracy  is \ncalculated  by  measuring the pixel  distance  between  the  centers  determined  manu(cid:173)\nally  on  the  original  image  and  as  the  center  of mass  in  the  image  obtained  after \napplication of the neural network.  Table  1 provides quantitative results. \nOn the whole) the two-step cascade of LLM networks yields for  399 out of 4 00 images \nan  activity  image precisely  centered  on  the  human hand.  Only in one  image)  the \nfirst  LLM  net missed the hand completely) due to a second  hand in the image that \ncould be clearly seen  in this view.  This image was excluded from further  processing \nand  from  the evaluation of the localization accuracy. \n\nCamera A \n\nCamera B \n\nPerson  A \nPerson  H \n\nPixel  deviatIOn \n\n0.8 \u00b1  1.2 \n1.3 \u00b1  1.4 \n\nNRMSE \n0.03  \u00b1 0.06 \n0.06 \u00b1 0.11 \n\nPixel deViatIOn \n\n0.8 \u00b1 2.2 \n2.2 \u00b1 2.8 \n\nNRMSE \n0.03 \u00b1 0.09 \n0.11  \u00b1 0.21 \n\nTable 1:  Estimation  error of the hand localization  on  the test set.  Absolute error in  pixels \nand normalized  error  for  both  persons  and  both  camera images. \n\n3.3  Recognition performance \nOne major problem in recognizing human pointing gestures is the variability of these \ngestures  and their measurement for  the acquisition of reliable training information. \nDifferent  persons follow  different  strategies where  and  how  to point  (fig.  2 (center) \nand  (right\u00bb.  Therefore)  we  calculate  this  information  indirectly.  The  person  is \ntold  to  point  at  a  certain  grid  position  with  known  world  coordinates.  From  the \ncamera images we  extract  the pixel  positions of the hand center  and map them to \nworld  coordinates  using the  PSOM  net  (PI  in  fig .  1).  Given these  coordinates the \nangles of the intended pointing vector with the basis vectors of the world coordinate \nsystem can  be  calculated trigonometrically.  These  angles form  the target vector for \nthe supervised  training of a  LLM  network  (D  in  fig.  1). \nAfter training) the output of the net is used to calculate the point where the pointing \nvector  intersects  the  table surface.  For  evaluation of the  network  performance  we \nmeasure the  Euclidian distance between this point and the actual grid point where \nthe person  intended to point at.  Fig. 3  (left)  shows the mean euclidean error MEE \nof the estimated target  position as  a function of the number of learning steps.  The \nerror  on the  training set  can  be  considerably  reduced)  whereas  on the test  set  the \nimprovement  stagnates  after  some  500  training  steps.  If we  perform  even  more \ntraining steps  the  performance  might actually suffer  from  overfitting.  The  graph \ncompares  training  and  test  results  achieved  on  images  obtained  by  two  different \nways of determining the hand  center.  The  \"manual\"  curves  show the performance \nthat  can  be  achieved  if the  Gabor  masks  are  manually centered  on the  hand.  For \nthe  \"neuronal))  curves)  the center of mass calculated in the fine-segmented  and post(cid:173)\nprocessed  subframe was  used.  This  allows  us  to study the  influence of the error  of \nthe segmentation and localization steps on the pointing recognition.  This influence \nis  rather small.  The MEE increases from  17  mm for  the optimal method to 19  mm \nfor  the neural  method)  which  is  hardly visible  in  practice. \nThe  curves  in  fig.  3  (center)  are  obtained  if we  apply  the  networks  to  images  of \nanother  person.  The  MEE  is  considerably  larger  but  a  detailed  analysis' shows \nthat part of this deviation is  due  to systematic differences  in the pointing strategy \nas  shown  in  fig.  2  (right).  Over  a  wide  range,  the  number  of nodes  used  for  the \nLLM  network  has  only  minor influence  on  the  performance.  While  obviously the \nperformance on the training set can be arbitrarily improved by spending more nodes, \nthe differences  in the MEE on the test set are negligible in a range of 5 to 15  nodes. \nUsing  more nodes  is  problematic as the training data consists of 50  examples only. \nIf not  indicated otherwise)  we  use  LLM  networks  with  10  nodes.  Further  results) \n\n\f908 \n\nE. LIITMANN. A. DREES. H.  RIITER \n\nFig.  2:  The  table  grid  points can  be reconstructed  according  to  the network output.  The \ntarget grid  is  dotted.  Reconstruction  of training  grid  (left)  and  test  grid  (center)  for  one \nperson,  and  of the  test grid  for  another person  (right). \n\nMER \n\nMEB on test oet of unknown perron \n\n30 \n\n:l~ \n\nm ..... aI,train-\n\nneuronal, train  -\nmanual, test  -\n\ne \n\nI~ \n\n10 \n\n---- ~--.---\n\n20  ~ 68 \n\u00a3  ~ \u00a3 \n\n66 \n64 \n62 \n60 \n58 \n56 \n\n~-\n\ne \n\n0 \n\n100 \n\n250  sao  1000  2SOO  SOOO \ntrain.., itHabonr \n\nn \n\n70  4 \n\n-~. \n\n100 \n\n:l~  sao  1000 \ntrairq IteratioN \n\n2SOO  SOOO \n\nFig.  3:  The euclidean  error of \nestimated  target point calcu-\nlated  using  the network out-\nput  depends  on  the  prepro-\ncessing (left),  and  the person \n(center). \n\ncomparing the pointing recognition based on only one of the camera images, indicate \nthat the method works better if the camera takes a lateral view rather than a frontal \nview . All evaluations were done for  both persons.  The performance was always very \nsimilar. \n4  Discussion \nWhile  we  begin  to  understand  many  properties  of neural  networks  at  the  single \nnetwork  level,  our  insight  into  principled  ways  of how  to  build  neural  systems  is \nstill  rather  limited .  Due  to  the  complexity  of this  task,  theoretical  progress  is \n(and  probably will  continue  to  be)  very  slow.  What we  can  do  in  the  mean time, \nhowever,  is  to  experiment  with  different  design  strategies  for  neural  systems  and \ntry to  \"evolve\"  useful  approaches  by  carefully chosen  case  studies. \nThe  current  work  is  an  effort  along  these  lines.  It  is  focused  on  a  challenging, \npractically important vision task with a  number of generic  features that  are shared \nwith vision tasks for  which  biological vision systems were  evolved. \nOne  important issue  is  how  to achieve  robustness  at the different  processing  levels \nof the  system.  There  are  only  very  limited  possibilities  to study  this  issue  in  si(cid:173)\nmulations, since practically nothing is  known  about the statistical properties of the \nvarious sources of error  that occur  when  dealing with real world data.  Thus,  a  real \nimplementation that  works  with  actual  data is  practically the  only  way  to  study \nthe robustness  issue  in  a  realistic fashion.  Therefore,  the  demonstrated integration \nof several  functional  modules that we  had  developed  previously  in  more restricted \nsettings  [7,  6]  was  a  non-trivial  test  of the  feasability  of  having  these  functions \ncooperate  in  a  larger,  modular  system.  It  also  gives  confidence  that  the  scaling \nproblem can  be  dealt with successfully  if we  apply modular neural  nets. \nA related  and equally important issue was the use of a processing strategy in which \nearlier  processing  stages  incrementally  restrict  the  search  space  for  the subsequent \nstages.  Thus, the responsibility for  achieving the goal is not centralized in any single \nmodule and subsequent  modules have always the chance to compensate for  limited \nerrors of earlier stages.  This appears to be a generally useful strategy for  achieving \n\n\fVisual Gesture-based Robot Guidance with a Modular Neural System \n\n909 \n\nrobustness  and for  cutting  computational costs  that is  related  to the  use  of \"focal \nattention\" , which is  clearly an important element of many biological vision systems. \nA  third important point is  the extensive  use  of learning to build the  essential  con(cid:173)\nstituent functions  of the system from  data examples.  We  are  not yet  able  to train \nthe  assembled system as  a whole.  Instead,  different  modules are trained separately \nand  are  integrated  only later.  Still,  the experience  gained with  assembling  a  com(cid:173)\nplex system via this  \"engineering-type\"  of approach will  be extremely valuable for \ngradually developing the  capability of crafting larger functional  building blocks by \nlearning methods. \nWe  conclude that carefully designed experiments with modular neural  systems that \nare  based  on  the  use  of real  world  data and  that  focus  on  similar tasks  for  which \nalso  biological neural  systems  were  evolved  can  make  a  significant  contribution  in \ntackling the challenge that lies ahead of us:  to develop  a reliable technology for  the \nconstruction of large-scale  artificial neural systems that can solve  complex tasks in \nreal  world environments. \nAcknowledgements \nWe want to thank Th.  Wengerek (robot control),  J.  Walter  (PSOM  implementation),  and \nP.  Ziemeck  (image  acquisition  software).  This work was supported  by  BMFT  Grant  No. \nITN9104AO. \n\nReferences \n[1]  T.  J.  Darell  and  A.  P.  Pentland.  Classifying  hand  gestures with  a  view-based  distri(cid:173)\n\nbuted  representation.  In  J .  D.  Cowan,  G.  Tesauro,  and  J.  Alspector,  editors,  Neural \nInformation  Processing Systems  6,  pages  945-952.  Morgan  Kaufman,  1994. \n\n[2]  J.  Davis and M.  Shah.  Recognizing  hand gestures.  In J.-O. Eklundh,  editor,  Computer \nVision  - ECCV  '94,  volume  800  of Lecture  Notes  in  Computer Science,  pages  331-\n340.  Springer-Verlag,  Berlin  Heidelberg  New York,  1994. \n\n[3]  R.A.  Jacobs,  M.1.  Jordan,  S.J.  Nowlan,  and  G.E.  Hinton.  Adaptive mixtures  of local \n\nexperts.  Neural  Computation,  3:79- 87,  1991. \n\n[4]  M.1.  Jordan and R.A.  Jacobs.  Hierarchical  mixtures of experts and the EM  algorithm. \n\nNeural  Computation,  6(2):181-214,  1994. \n\n[5]  F.  Kummert,  E.  Littmann,  A.  Meyering,  S.  Posch,  H.  Ritter,  and  G.  Sagerer.  A \nIn \n\nhybrid  approach  to  signal  interpretation  using  neural  and  semantic  networks. \nMustererkennung 1993,  pages  245-252.  Springer,  1993. \n\n[6]  E.  Littmann,  A.  Drees,  and  H.  Ritter.  Neural recognition  of human pointing gestures \n\nin  real  images.  Submitted  to  Neural Processing Letters,  1996. \n\n[7]  E.  Littmann  and  H.  Ritter.  Neural  and  statistical  methods  for  adaptive  color  seg(cid:173)\nIn  G.  Sagerer,  S.  Posch,  and  F.  Kummert,  editors, \n\nmentation  -\nMustererkennung 1995, pages  84-93.  Springer-Verlag,  Heidelberg,  1995. \n\na  comparison. \n\n[8]  C.  Maggioni.  A  novel  device  for  using  the  hand  as  a  human-computer  interface.  In \nProceedings HC1'93 - Human  Control Interface, Loughborough,  Great Britain,  1993. \n[9]  A.  Meyering  and  H.  Ritter.  Learning  3D  shape  perception  with local  linear  maps.  In \n\nProc.  of the lJCNN, volume  IV,  pages  432-436,  Baltimore,  MD,  1992. \n\n[10]  Steven  J.  Nowlan  and  John  C.  Platt.  A  convolutional  neural  network  hand  tracker. \n\nIn  Neural  Information Processing Systems  7.  Morgan  Kaufman  Publishers,  1995. \n\n[11]  H.  Ritter.  Parametrized self-organizing  maps for vision  learning tasks.  In  P.  Morasso, \n\neditor,  ICANN  '94.  Springer-Verlag,  Berlin  Heidelberg  New York,  1994. \n\n[12]  K.  Viiiina.nen  and  K.  Bohm.  Gesture driven  interaction  as  a  human  factor  in  virtual \nenvironments - an  approach  with neural  networks.  In R.  Earnshaw,  M.  Gigante,  and \nH.  Jones,  editors,  Virtual reality  systems, pages  93-106.  Academic  Press,  1993. \n\n[13]  J.  Walter  and  H.  Ritter.  Rapid  learning  with  parametrized  self-organizing  maps. \n\nNeural  Computing,  1995.  Submitted. \n\n[14]  T.  G.  Zimmermann,  J.  Lanier,  C.  Blanchard,  S.  Bryson,  and  Y.  Harvill.  A  hand \n\ngesture  interface  device.  In  Proc.  CHI+GI,  pages 189-192,  1987. \n\n\f", "award": [], "sourceid": 1074, "authors": [{"given_name": "Enno", "family_name": "Littmann", "institution": null}, {"given_name": "Andrea", "family_name": "Drees", "institution": null}, {"given_name": "Helge", "family_name": "Ritter", "institution": null}]}