00001 /*!@file Channels/DirectionOpticalFlowChannel.C */ 00002 // //////////////////////////////////////////////////////////////////// // 00003 // The iLab Neuromorphic Vision C++ Toolkit - Copyright (C) 2000-2005 // 00004 // by the University of Southern California (USC) and the iLab at USC. // 00005 // See http://iLab.usc.edu for information about this project. // 00006 // //////////////////////////////////////////////////////////////////// // 00007 // Major portions of the iLab Neuromorphic Vision Toolkit are protected // 00008 // under the U.S. patent ``Computation of Intrinsic Perceptual Saliency // 00009 // in Visual Environments, and Applications'' by Christof Koch and // 00010 // Laurent Itti, California Institute of Technology, 2001 (patent // 00011 // pending; application number 09/912,225 filed July 23, 2001; see // 00012 // http://pair.uspto.gov/cgi-bin/final/home.pl for current status). // 00013 // //////////////////////////////////////////////////////////////////// // 00014 // This file is part of the iLab Neuromorphic Vision C++ Toolkit. // 00015 // // 00016 // The iLab Neuromorphic Vision C++ Toolkit is free software; you can // 00017 // redistribute it and/or modify it under the terms of the GNU General // 00018 // Public License as published by the Free Software Foundation; either // 00019 // version 2 of the License, or (at your option) any later version. // 00020 // // 00021 // The iLab Neuromorphic Vision C++ Toolkit is distributed in the hope // 00022 // that it will be useful, but WITHOUT ANY WARRANTY; without even the // 00023 // implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR // 00024 // PURPOSE. See the GNU General Public License for more details. // 00025 // // 00026 // You should have received a copy of the GNU General Public License // 00027 // along with the iLab Neuromorphic Vision C++ Toolkit; if not, write // 00028 // to the Free Software Foundation, Inc., 59 Temple Place, Suite 330, // 00029 // Boston, MA 02111-1307 USA. // 00030 // //////////////////////////////////////////////////////////////////// // 00031 // 00032 // Primary maintainer for this file: 00033 // $HeadURL: svn://isvn.usc.edu/software/invt/trunk/saliency/src/Channels/DirectionOpticalFlowChannel.C $ 00034 // $Id:$ 00035 // 00036 00037 #ifndef DIRECTION_OPTICALFLOW_CHANNEL_C_DEFINED 00038 #define DIRECTION_OPTICALFLOW_CHANNEL_C_DEFINED 00039 00040 #include "Channels/DirectionOpticalFlowChannel.H" 00041 00042 #include "rutz/trace.h" 00043 #include "Image/Image.H" 00044 #include "Raster/Raster.H" 00045 00046 #define OF_WINDOW_SIZE 4 00047 #define OF_VALUE_RADIUS 1 00048 00049 // ###################################################################### 00050 // DirectionOpticalFlowChannel member definitions: 00051 // ###################################################################### 00052 00053 // ###################################################################### 00054 DirectionOpticalFlowChannel::DirectionOpticalFlowChannel 00055 (OptionManager& mgr, 00056 const uint dirIndex, const double direction, 00057 const OpticalFlowType type): 00058 SingleChannel(mgr, "", "", MOTIONSPATIOTEMPORAL, 00059 rutz::make_shared(new GaussianPyrBuilder<float>(1))), // no actual pyrbuilder 00060 itsDirIndex("DirectionChannelDirectionIndex", this, dirIndex), 00061 itsDirection("DirectionChannelDirection", this, direction), 00062 itsOpticalFlowType(type) 00063 { 00064 GVX_TRACE(__PRETTY_FUNCTION__); 00065 00066 setDescriptiveName(sformat("DirectionSpeed(%d)", int(direction))); 00067 setTagName(sformat("dir_%d", dirIndex)); 00068 } 00069 00070 // ###################################################################### 00071 void DirectionOpticalFlowChannel::start1() 00072 { 00073 GVX_TRACE(__PRETTY_FUNCTION__); 00074 itsLevelSpec.setVal(LevelSpec(0,0,0,0,4) ); // only 1 level 00075 SingleChannel::start1(); 00076 } 00077 00078 // ###################################################################### 00079 void DirectionOpticalFlowChannel::start2() 00080 { 00081 GVX_TRACE(__PRETTY_FUNCTION__); 00082 00083 } 00084 00085 // ###################################################################### 00086 DirectionOpticalFlowChannel::~DirectionOpticalFlowChannel() 00087 { 00088 GVX_TRACE(__PRETTY_FUNCTION__); 00089 } 00090 00091 // ###################################################################### 00092 bool DirectionOpticalFlowChannel::outputAvailable() const 00093 { 00094 return itsDirectionalOpticalFlow.initialized(); 00095 } 00096 00097 // ###################################################################### 00098 //! different ways to input different optical flow algorithms 00099 void DirectionOpticalFlowChannel::setLucasKanadeOpticalFlow 00100 (rutz::shared_ptr<OpticalFlow> flow) 00101 { 00102 ASSERT(itsOpticalFlowType == LucasKanade); 00103 itsOpticalFlow = flow; 00104 } 00105 00106 // ###################################################################### 00107 void DirectionOpticalFlowChannel::doInput(const InputFrame& inframe) 00108 { 00109 GVX_TRACE(__PRETTY_FUNCTION__); 00110 00111 if (!this->started()) 00112 CLFATAL("must be start()-ed before using receiving any input"); 00113 00114 // so that the visual cortex think we can produce a conspicuity map 00115 // FIXXX: really stupid hack 00116 SingleChannel::doInput(inframe); // will store the new pyramid 00117 00118 // check if the optical flow input is already provided 00119 switch(itsOpticalFlowType) 00120 { 00121 case LucasKanade: 00122 if(itsOpticalFlow.is_invalid()) 00123 LFATAL("need to provide optical flow. " 00124 "Use setLucasKanadeOpticalFlow(flow)"); 00125 computeDirectionalOpticalFlow(); 00126 break; 00127 case HornSchunck: 00128 break; 00129 default: LERROR("Unknown optical flow mode"); 00130 } 00131 } 00132 00133 // ###################################################################### 00134 void DirectionOpticalFlowChannel::computeDirectionalOpticalFlow() 00135 { 00136 Dims d = itsOpticalFlow->getImageDims(); 00137 itsDirectionalOpticalFlow = 00138 Image<float>(d.w()/ OF_WINDOW_SIZE, d.h()/ OF_WINDOW_SIZE, ZEROS); 00139 00140 // go through each flowVector using std::vector 00141 // FIXXX: for now. We know that Lucas Kanade produces sparse vectors. 00142 00143 std::vector<rutz::shared_ptr<FlowVector> > flow = 00144 itsOpticalFlow->getFlowVectors(); 00145 00146 for(uint i = 0; i < flow.size(); i++) 00147 { 00148 Point2D<int> pt1((int)flow[i]->p1.i, (int)flow[i]->p1.j); 00149 Point2D<int> pt2((int)flow[i]->p2.i, (int)flow[i]->p2.j); 00150 00151 float ang = flow[i]->angle; 00152 //float mag = flow[i]->mag; 00153 float fval = flow[i]->val; 00154 00155 // get the response value of the 00156 float rv = getResponseValue(ang); 00157 computeDirectionalOpticalFlow(pt1.i, pt1.j, rv*fval); 00158 00159 //LINFO("[%3d %3d]: ang: %f mag: %f fval: %f ---> %f", 00160 // pt1.i, pt1.j, ang, mag, fval, rv*fval); 00161 } 00162 } 00163 00164 // ###################################################################### 00165 float DirectionOpticalFlowChannel::getResponseValue(float angle) 00166 { 00167 float pangle = itsDirection.getVal(); 00168 00169 // map to to 0 - 360 deg 00170 float nangle = fmod(angle +360.0, 360.0); 00171 00172 // find difference in direction 00173 float diff1 = fabs(nangle - pangle); 00174 float diff2 = 360.0 - diff1; 00175 float diff = diff1; if(diff1 > diff2) diff = diff2; 00176 00177 // we are going to focus to between +/- 30 deg diff 00178 float val = 0.0; 00179 float stdang = 15.0; 00180 00181 // difference between 0 and 15 degrees 00182 if(diff >= 0 && diff < stdang) 00183 val = (stdang - diff)/stdang * 0.2 + 0.8; 00184 00185 // difference between 15 and 30 degrees 00186 else if(diff >= stdang && diff < 2*stdang) 00187 val = (2*stdang - diff)/stdang * 0.8 + 0.0; 00188 00189 // difference farther than 30 degrees 00190 else 00191 val = 0.0; 00192 00193 return val; 00194 } 00195 00196 // ###################################################################### 00197 void DirectionOpticalFlowChannel::computeDirectionalOpticalFlow 00198 (uint i, uint j, float val) 00199 { 00200 // scale down to the actual directional flow map 00201 uint ii = i/OF_WINDOW_SIZE; 00202 uint jj = j/OF_WINDOW_SIZE; 00203 00204 00205 //itsDirectionalOpticalFlow.setVal(ii, jj, val); 00206 00207 uint ri = OF_VALUE_RADIUS; 00208 uint rj = OF_VALUE_RADIUS; 00209 00210 uint width = itsDirectionalOpticalFlow.getWidth(); 00211 uint height = itsDirectionalOpticalFlow.getHeight(); 00212 00213 uint lm = 0; if(ii > ri-1) lm = ii - ri; 00214 uint rm = width-1; if(ii < width-1-ri) rm = ii + ri; 00215 uint tm = 0; if(jj > rj-1) tm = jj - rj; 00216 uint bm = height-1; if(jj < height-1-ri) bm = jj + rj; 00217 00218 // set the values 00219 float max = double (ri*2); 00220 for(uint bi = lm; bi <= rm; bi++) 00221 { 00222 for(uint bj = tm; bj <= bm; bj++) 00223 { 00224 // FIX: pre-compute the weights later 00225 float di = float(bi) - float(ii); 00226 float dj = float(bj) - float(jj); 00227 float w = 1.0 - (sqrt(di*di + dj*dj))/max; 00228 float ov = itsDirectionalOpticalFlow.getVal(bi, bj); 00229 itsDirectionalOpticalFlow.setVal(bi, bj, ov+w*val); 00230 //LINFO("%d %d: %f", bi, bj, w*val); 00231 } 00232 } 00233 } 00234 00235 // ###################################################################### 00236 Image<float> DirectionOpticalFlowChannel::getDirectionalOpticalFlow() 00237 { 00238 return itsDirectionalOpticalFlow; 00239 } 00240 00241 // ###################################################################### 00242 /* So things look consistent in everyone's emacs... */ 00243 /* Local Variables: */ 00244 /* indent-tabs-mode: nil */ 00245 /* End: */ 00246 00247 #endif // DIRECTION_OPTICALFLOW_CHANNEL_C_DEFINED