bgfg_gmg.cpp 18.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
/*M///////////////////////////////////////////////////////////////////////////////////////
//
//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
//  By downloading, copying, installing or using the software you agree to this license.
//  If you do not agree to this license, do not download, install,
//  copy or use the software.
//
//
//                          License Agreement
//                For Open Source Computer Vision Library
//
// Copyright (C) 2000, Intel Corporation, all rights reserved.
// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
//   * Redistribution's of source code must retain the above copyright notice,
//     this list of conditions and the following disclaimer.
//
//   * Redistribution's in binary form must reproduce the above copyright notice,
//     this list of conditions and the following disclaimer in the documentation
//     and/or other materials provided with the distribution.
//
//   * The name of the copyright holders may not be used to endorse or promote products
//     derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/

/*
44
 * This class implements a particular BackgroundSubtraction algorithm described in "Visual Tracking of Human Visitors under
45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
 * Variable-Lighting Conditions for a Responsive Audio Art Installation," A. Godbehere,
 * A. Matsukawa, K. Goldberg, American Control Conference, Montreal, June 2012.
 *
 * Prepared and integrated by Andrew B. Godbehere.
 */

#include "precomp.hpp"
#include "opencv2/core/utility.hpp"
#include <limits>

namespace cv
{
namespace bgsegm
{

60
class BackgroundSubtractorGMGImpl CV_FINAL : public BackgroundSubtractorGMG
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
{
public:
    BackgroundSubtractorGMGImpl()
    {
        /*
         * Default Parameter Values. Override with algorithm "set" method.
         */
        maxFeatures = 64;
        learningRate = 0.025;
        numInitializationFrames = 120;
        quantizationLevels = 16;
        backgroundPrior = 0.8;
        decisionThreshold = 0.8;
        smoothingRadius = 7;
        updateBackgroundModel = true;
        minVal_ = maxVal_ = 0;
        name_ = "BackgroundSubtractor.GMG";
    }

    ~BackgroundSubtractorGMGImpl()
    {
    }

    /**
     * Validate parameters and set up data structures for appropriate image size.
     * Must call before running on data.
     * @param frameSize input frame size
     * @param min       minimum value taken on by pixels in image sequence. Usually 0
     * @param max       maximum value taken on by pixels in image sequence. e.g. 1.0 or 255
     */
    void initialize(Size frameSize, double minVal, double maxVal);

    /**
     * Performs single-frame background subtraction and builds up a statistical background image
     * model.
     * @param image Input image
     * @param fgmask Output mask image representing foreground and background pixels
     */
99
    virtual void apply(InputArray image, OutputArray fgmask, double learningRate=-1.0) CV_OVERRIDE;
100 101 102 103 104 105

    /**
     * Releases all inner buffers.
     */
    void release();

106 107
    virtual int getMaxFeatures() const CV_OVERRIDE { return maxFeatures; }
    virtual void setMaxFeatures(int _maxFeatures) CV_OVERRIDE { maxFeatures = _maxFeatures; }
108

109 110
    virtual double getDefaultLearningRate() const CV_OVERRIDE { return learningRate; }
    virtual void setDefaultLearningRate(double lr) CV_OVERRIDE { learningRate = lr; }
111

112 113
    virtual int getNumFrames() const CV_OVERRIDE { return numInitializationFrames; }
    virtual void setNumFrames(int nframes) CV_OVERRIDE { numInitializationFrames = nframes; }
114

115 116
    virtual int getQuantizationLevels() const CV_OVERRIDE { return quantizationLevels; }
    virtual void setQuantizationLevels(int nlevels) CV_OVERRIDE { quantizationLevels = nlevels; }
117

118 119
    virtual double getBackgroundPrior() const CV_OVERRIDE { return backgroundPrior; }
    virtual void setBackgroundPrior(double bgprior) CV_OVERRIDE { backgroundPrior = bgprior; }
120

121 122
    virtual int getSmoothingRadius() const CV_OVERRIDE { return smoothingRadius; }
    virtual void setSmoothingRadius(int radius) CV_OVERRIDE { smoothingRadius = radius; }
123

124 125
    virtual double getDecisionThreshold() const CV_OVERRIDE { return decisionThreshold; }
    virtual void setDecisionThreshold(double thresh) CV_OVERRIDE { decisionThreshold = thresh; }
126

127 128
    virtual bool getUpdateBackgroundModel() const CV_OVERRIDE { return updateBackgroundModel; }
    virtual void setUpdateBackgroundModel(bool update) CV_OVERRIDE { updateBackgroundModel = update; }
129

130 131
    virtual double getMinVal() const CV_OVERRIDE { return minVal_; }
    virtual void setMinVal(double val) CV_OVERRIDE { minVal_ = val; }
132

133 134
    virtual double getMaxVal() const CV_OVERRIDE { return maxVal_; }
    virtual void setMaxVal(double val) CV_OVERRIDE { maxVal_ = val; }
135

136
    virtual void getBackgroundImage(OutputArray backgroundImage) const CV_OVERRIDE
137 138 139 140
    {
        backgroundImage.release();
    }

141
    virtual void write(FileStorage& fs) const CV_OVERRIDE
142 143 144 145 146 147 148 149 150 151 152 153 154
    {
        fs << "name" << name_
        << "maxFeatures" << maxFeatures
        << "defaultLearningRate" << learningRate
        << "numFrames" << numInitializationFrames
        << "quantizationLevels" << quantizationLevels
        << "backgroundPrior" << backgroundPrior
        << "decisionThreshold" << decisionThreshold
        << "smoothingRadius" << smoothingRadius
        << "updateBackgroundModel" << (int)updateBackgroundModel;
        // we do not save minVal_ & maxVal_, since they depend on the image type.
    }

155
    virtual void read(const FileNode& fn) CV_OVERRIDE
156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
    {
        CV_Assert( (String)fn["name"] == name_ );
        maxFeatures = (int)fn["maxFeatures"];
        learningRate = (double)fn["defaultLearningRate"];
        numInitializationFrames = (int)fn["numFrames"];
        quantizationLevels = (int)fn["quantizationLevels"];
        backgroundPrior = (double)fn["backgroundPrior"];
        smoothingRadius = (int)fn["smoothingRadius"];
        decisionThreshold = (double)fn["decisionThreshold"];
        updateBackgroundModel = (int)fn["updateBackgroundModel"] != 0;
        minVal_ = maxVal_ = 0;
        frameSize_ = Size();
    }

    //! Total number of distinct colors to maintain in histogram.
    int     maxFeatures;
    //! Set between 0.0 and 1.0, determines how quickly features are "forgotten" from histograms.
    double  learningRate;
    //! Number of frames of video to use to initialize histograms.
    int     numInitializationFrames;
    //! Number of discrete levels in each channel to be used in histograms.
    int     quantizationLevels;
    //! Prior probability that any given pixel is a background pixel. A sensitivity parameter.
    double  backgroundPrior;
    //! Value above which pixel is determined to be FG.
    double  decisionThreshold;
    //! Smoothing radius, in pixels, for cleaning up FG image.
    int     smoothingRadius;
    //! Perform background model update
    bool updateBackgroundModel;

private:
    double maxVal_;
    double minVal_;

    Size frameSize_;
    int frameNum_;

    String name_;

    Mat_<int> nfeatures_;
197
    Mat_<int> colors_;
198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
    Mat_<float> weights_;
};


void BackgroundSubtractorGMGImpl::initialize(Size frameSize, double minVal, double maxVal)
{
    CV_Assert(minVal < maxVal);
    CV_Assert(maxFeatures > 0);
    CV_Assert(learningRate >= 0.0 && learningRate <= 1.0);
    CV_Assert(numInitializationFrames >= 1);
    CV_Assert(quantizationLevels >= 1 && quantizationLevels <= 255);
    CV_Assert(backgroundPrior >= 0.0 && backgroundPrior <= 1.0);

    minVal_ = minVal;
    maxVal_ = maxVal;

    frameSize_ = frameSize;
    frameNum_ = 0;

    nfeatures_.create(frameSize_);
    colors_.create(frameSize_.area(), maxFeatures);
    weights_.create(frameSize_.area(), maxFeatures);

    nfeatures_.setTo(Scalar::all(0));
}

224
static float findFeature(int color, const int* colors, const float* weights, int nfeatures)
225 226 227 228 229 230 231 232 233 234 235
{
    for (int i = 0; i < nfeatures; ++i)
    {
        if (color == colors[i])
            return weights[i];
    }

    // not in histogram, so return 0.
    return 0.0f;
}

236
static void normalizeHistogram(float* weights, int nfeatures)
237 238 239 240 241 242 243 244 245 246 247 248
{
    float total = 0.0f;
    for (int i = 0; i < nfeatures; ++i)
        total += weights[i];

    if (total != 0.0f)
    {
        for (int i = 0; i < nfeatures; ++i)
            weights[i] /= total;
    }
}

249
static bool insertFeature(int color, float weight, int* colors, float* weights, int& nfeatures, int maxFeatures)
250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266
{
    int idx = -1;
    for (int i = 0; i < nfeatures; ++i)
    {
        if (color == colors[i])
        {
            // feature in histogram
            weight += weights[i];
            idx = i;
            break;
        }
    }

    if (idx >= 0)
    {
        // move feature to beginning of list

267
        ::memmove(colors + 1, colors, idx * sizeof(int));
268 269 270 271 272 273 274 275 276
        ::memmove(weights + 1, weights, idx * sizeof(float));

        colors[0] = color;
        weights[0] = weight;
    }
    else if (nfeatures == maxFeatures)
    {
        // discard oldest feature

277
        ::memmove(colors + 1, colors, (nfeatures - 1) * sizeof(int));
278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297
        ::memmove(weights + 1, weights, (nfeatures - 1) * sizeof(float));

        colors[0] = color;
        weights[0] = weight;
    }
    else
    {
        colors[nfeatures] = color;
        weights[nfeatures] = weight;

        ++nfeatures;

        return true;
    }

    return false;
}

template <typename T> struct Quantization
{
298
    static int apply(const void* src_, int x, int cn, double minVal, double maxVal, int quantizationLevels)
299 300 301 302 303 304 305 306 307 308 309 310 311 312 313
    {
        const T* src = static_cast<const T*>(src_);
        src += x * cn;

        unsigned int res = 0;
        for (int i = 0, shift = 0; i < cn; ++i, ++src, shift += 8)
            res |= static_cast<int>((*src - minVal) * quantizationLevels / (maxVal - minVal)) << shift;

        return res;
    }
};

class GMG_LoopBody : public ParallelLoopBody
{
public:
314
    GMG_LoopBody(const Mat& frame, const Mat& fgmask, const Mat_<int>& nfeatures, const Mat_<int>& colors, const Mat_<float>& weights,
315 316 317 318 319 320 321 322 323
                 int maxFeatures, double learningRate, int numInitializationFrames, int quantizationLevels, double backgroundPrior, double decisionThreshold,
                 double maxVal, double minVal, int frameNum, bool updateBackgroundModel) :
        frame_(frame), fgmask_(fgmask), nfeatures_(nfeatures), colors_(colors), weights_(weights),
        maxFeatures_(maxFeatures), learningRate_(learningRate), numInitializationFrames_(numInitializationFrames), quantizationLevels_(quantizationLevels),
        backgroundPrior_(backgroundPrior), decisionThreshold_(decisionThreshold), updateBackgroundModel_(updateBackgroundModel),
        maxVal_(maxVal), minVal_(minVal), frameNum_(frameNum)
    {
    }

324
    void operator() (const Range& range) const CV_OVERRIDE;
325 326 327 328 329 330 331

private:
    Mat frame_;

    mutable Mat_<uchar> fgmask_;

    mutable Mat_<int> nfeatures_;
332
    mutable Mat_<int> colors_;
333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349
    mutable Mat_<float> weights_;

    int     maxFeatures_;
    double  learningRate_;
    int     numInitializationFrames_;
    int     quantizationLevels_;
    double  backgroundPrior_;
    double  decisionThreshold_;
    bool updateBackgroundModel_;

    double maxVal_;
    double minVal_;
    int frameNum_;
};

void GMG_LoopBody::operator() (const Range& range) const
{
350
    typedef int (*func_t)(const void* src_, int x, int cn, double minVal, double maxVal, int quantizationLevels);
351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375
    static const func_t funcs[] =
    {
        Quantization<uchar>::apply,
        Quantization<schar>::apply,
        Quantization<ushort>::apply,
        Quantization<short>::apply,
        Quantization<int>::apply,
        Quantization<float>::apply,
        Quantization<double>::apply
    };

    const func_t func = funcs[frame_.depth()];
    CV_Assert(func != 0);

    const int cn = frame_.channels();

    for (int y = range.start, featureIdx = y * frame_.cols; y < range.end; ++y)
    {
        const uchar* frame_row = frame_.ptr(y);
        int* nfeatures_row = nfeatures_[y];
        uchar* fgmask_row = fgmask_[y];

        for (int x = 0; x < frame_.cols; ++x, ++featureIdx)
        {
            int nfeatures = nfeatures_row[x];
376
            int* colors = colors_[featureIdx];
377 378
            float* weights = weights_[featureIdx];

379
            int newFeatureColor = func(frame_row, x, cn, minVal_, maxVal_, quantizationLevels_);
380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459

            bool isForeground = false;

            if (frameNum_ >= numInitializationFrames_)
            {
                // typical operation

                const double weight = findFeature(newFeatureColor, colors, weights, nfeatures);

                // see Godbehere, Matsukawa, Goldberg (2012) for reasoning behind this implementation of Bayes rule
                const double posterior = (weight * backgroundPrior_) / (weight * backgroundPrior_ + (1.0 - weight) * (1.0 - backgroundPrior_));

                isForeground = ((1.0 - posterior) > decisionThreshold_);

                // update histogram.

                if (updateBackgroundModel_)
                {
                    for (int i = 0; i < nfeatures; ++i)
                        weights[i] *= (float)(1.0f - learningRate_);

                    bool inserted = insertFeature(newFeatureColor, (float)learningRate_, colors, weights, nfeatures, maxFeatures_);

                    if (inserted)
                    {
                        normalizeHistogram(weights, nfeatures);
                        nfeatures_row[x] = nfeatures;
                    }
                }
            }
            else if (updateBackgroundModel_)
            {
                // training-mode update

                insertFeature(newFeatureColor, 1.0f, colors, weights, nfeatures, maxFeatures_);

                if (frameNum_ == numInitializationFrames_ - 1)
                    normalizeHistogram(weights, nfeatures);
            }

            fgmask_row[x] = (uchar)(-(schar)isForeground);
        }
    }
}

void BackgroundSubtractorGMGImpl::apply(InputArray _frame, OutputArray _fgmask, double newLearningRate)
{
    Mat frame = _frame.getMat();

    CV_Assert(frame.depth() == CV_8U || frame.depth() == CV_16U || frame.depth() == CV_32F);
    CV_Assert(frame.channels() == 1 || frame.channels() == 3 || frame.channels() == 4);

    if (newLearningRate != -1.0)
    {
        CV_Assert(newLearningRate >= 0.0 && newLearningRate <= 1.0);
        learningRate = newLearningRate;
    }

    if (frame.size() != frameSize_)
    {
        double minval = minVal_;
        double maxval = maxVal_;
        if( minVal_ == 0 && maxVal_ == 0 )
        {
            minval = 0;
            maxval = frame.depth() == CV_8U ? 255.0 : frame.depth() == CV_16U ? std::numeric_limits<ushort>::max() : 1.0;
        }
        initialize(frame.size(), minval, maxval);
    }

    _fgmask.create(frameSize_, CV_8UC1);
    Mat fgmask = _fgmask.getMat();

    GMG_LoopBody body(frame, fgmask, nfeatures_, colors_, weights_,
                      maxFeatures, learningRate, numInitializationFrames, quantizationLevels, backgroundPrior, decisionThreshold,
                      maxVal_, minVal_, frameNum_, updateBackgroundModel);
    parallel_for_(Range(0, frame.rows), body, frame.total()/(double)(1<<16));

    if (smoothingRadius > 0)
    {
460
        medianBlur(fgmask, fgmask, smoothingRadius);
461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515
    }

    // keep track of how many frames we have processed
    ++frameNum_;
}

void BackgroundSubtractorGMGImpl::release()
{
    frameSize_ = Size();

    nfeatures_.release();
    colors_.release();
    weights_.release();
}


Ptr<BackgroundSubtractorGMG> createBackgroundSubtractorGMG(int initializationFrames, double decisionThreshold)
{
    Ptr<BackgroundSubtractorGMG> bgfg = makePtr<BackgroundSubtractorGMGImpl>();
    bgfg->setNumFrames(initializationFrames);
    bgfg->setDecisionThreshold(decisionThreshold);

    return bgfg;
}

/*
 ///////////////////////////////////////////////////////////////////////////////////////////////////////////

 CV_INIT_ALGORITHM(BackgroundSubtractorGMG, "BackgroundSubtractor.GMG",
 obj.info()->addParam(obj, "maxFeatures", obj.maxFeatures,false,0,0,
 "Maximum number of features to store in histogram. Harsh enforcement of sparsity constraint.");
 obj.info()->addParam(obj, "learningRate", obj.learningRate,false,0,0,
 "Adaptation rate of histogram. Close to 1, slow adaptation. Close to 0, fast adaptation, features forgotten quickly.");
 obj.info()->addParam(obj, "initializationFrames", obj.numInitializationFrames,false,0,0,
 "Number of frames to use to initialize histograms of pixels.");
 obj.info()->addParam(obj, "quantizationLevels", obj.quantizationLevels,false,0,0,
 "Number of discrete colors to be used in histograms. Up-front quantization.");
 obj.info()->addParam(obj, "backgroundPrior", obj.backgroundPrior,false,0,0,
 "Prior probability that each individual pixel is a background pixel.");
 obj.info()->addParam(obj, "smoothingRadius", obj.smoothingRadius,false,0,0,
 "Radius of smoothing kernel to filter noise from FG mask image.");
 obj.info()->addParam(obj, "decisionThreshold", obj.decisionThreshold,false,0,0,
 "Threshold for FG decision rule. Pixel is FG if posterior probability exceeds threshold.");
 obj.info()->addParam(obj, "updateBackgroundModel", obj.updateBackgroundModel,false,0,0,
 "Perform background model update.");
 obj.info()->addParam(obj, "minVal", obj.minVal_,false,0,0,
 "Minimum of the value range (mostly for regression testing)");
 obj.info()->addParam(obj, "maxVal", obj.maxVal_,false,0,0,
 "Maximum of the value range (mostly for regression testing)");
 );
*/

}
}