Pedestrian-Demo

Prev Tutorial: Font-Demo
Next Tutorial: Optflow-Demo

Original author	Amir Hassan (kallaballa) amir@.nosp@m.viel.nosp@m.-zu.o.nosp@m.rg
Compatibility	OpenCV >= 4.7

Pedestrian detection using HOG with a linear SVM, non-maximal suppression and tracking using KCF. Uses nanovg for rendering (OpenGL), detects using a linear SVM (OpenCV), filters resuls using NMS and tracks using KCF.

Start CaptureFullscreen

Downloading...

// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
// Copyright Amir Hassan (kallaballa) <amir@viel-zu.org>
 
#include <opencv2/v4d/v4d.hpp>
#include <opencv2/tracking.hpp>
#include <opencv2/objdetect.hpp>
 
#include <string>
 
using std::cerr;
using std::endl;
using std::vector;
using std::string;
 
/* Demo parameters */
#ifndef __EMSCRIPTEN__
constexpr long unsigned int WIDTH = 1280;
constexpr long unsigned int HEIGHT = 720;
#else
constexpr long unsigned int WIDTH = 960;
constexpr long unsigned int HEIGHT = 960;
#endif
const unsigned long DIAG = hypot(double(WIDTH), double(HEIGHT));
constexpr unsigned int DOWNSIZE_WIDTH = 640;
constexpr unsigned int DOWNSIZE_HEIGHT = 360;
constexpr double WIDTH_SCALE = double(WIDTH) / DOWNSIZE_WIDTH;
constexpr double HEIGHT_SCALE = double(HEIGHT) / DOWNSIZE_HEIGHT;
constexpr bool OFFSCREEN = false;
#ifndef __EMSCRIPTEN__
constexpr const char* OUTPUT_FILENAME = "pedestrian-demo.mkv";
#endif
const int BLUR_KERNEL_SIZE = std::max(int(DIAG / 200 % 2 == 0 ? DIAG / 200 + 1 : DIAG / 200), 1);
 
 
//adapted from cv::dnn_objdetect::InferBbox
static inline bool pair_comparator(std::pair<double, size_t> l1, std::pair<double, size_t> l2) {
    return l1.first > l2.first;
}
 
//adapted from cv::dnn_objdetect::InferBbox
static void intersection_over_union(std::vector<std::vector<double> > *boxes, std::vector<double> *base_box, std::vector<double> *iou) {
    double g_xmin = (*base_box)[0];
    double g_ymin = (*base_box)[1];
    double g_xmax = (*base_box)[2];
    double g_ymax = (*base_box)[3];
    double base_box_w = g_xmax - g_xmin;
    double base_box_h = g_ymax - g_ymin;
    for (size_t b = 0; b < (*boxes).size(); ++b) {
        double xmin = std::max((*boxes)[b][0], g_xmin);
        double ymin = std::max((*boxes)[b][1], g_ymin);
        double xmax = std::min((*boxes)[b][2], g_xmax);
        double ymax = std::min((*boxes)[b][3], g_ymax);
 
        // Intersection
        double w = std::max(static_cast<double>(0.0), xmax - xmin);
        double h = std::max(static_cast<double>(0.0), ymax - ymin);
        // Union
        double test_box_w = (*boxes)[b][2] - (*boxes)[b][0];
        double test_box_h = (*boxes)[b][3] - (*boxes)[b][1];
 
        double inter_ = w * h;
        double union_ = test_box_h * test_box_w + base_box_h * base_box_w - inter_;
        (*iou)[b] = inter_ / (union_ + 1e-7);
    }
}
 
//adapted from cv::dnn_objdetect::InferBbox
static std::vector<bool> non_maximal_suppression(std::vector<std::vector<double> > *boxes, std::vector<double> *probs, const double threshold = 0.1) {
    std::vector<bool> keep(((*probs).size()));
    std::fill(keep.begin(), keep.end(), true);
    std::vector<size_t> prob_args_sorted((*probs).size());
 
    std::vector<std::pair<double, size_t> > temp_sort((*probs).size());
    for (size_t tidx = 0; tidx < (*probs).size(); ++tidx) {
        temp_sort[tidx] = std::make_pair((*probs)[tidx], static_cast<size_t>(tidx));
    }
    std::sort(temp_sort.begin(), temp_sort.end(), pair_comparator);
 
    for (size_t idx = 0; idx < temp_sort.size(); ++idx) {
        prob_args_sorted[idx] = temp_sort[idx].second;
    }
 
    for (std::vector<size_t>::iterator itr = prob_args_sorted.begin(); itr != prob_args_sorted.end() - 1; ++itr) {
        size_t idx = itr - prob_args_sorted.begin();
        std::vector<double> iou_(prob_args_sorted.size() - idx - 1);
        std::vector<std::vector<double> > temp_boxes(iou_.size());
        for (size_t bb = 0; bb < temp_boxes.size(); ++bb) {
            std::vector<double> temp_box(4);
            for (size_t b = 0; b < 4; ++b) {
                temp_box[b] = (*boxes)[prob_args_sorted[idx + bb + 1]][b];
            }
            temp_boxes[bb] = temp_box;
        }
        intersection_over_union(&temp_boxes, &(*boxes)[prob_args_sorted[idx]], &iou_);
        for (std::vector<double>::iterator _itr = iou_.begin(); _itr != iou_.end(); ++_itr) {
            size_t iou_idx = _itr - iou_.begin();
            if (*_itr > threshold) {
                keep[prob_args_sorted[idx + iou_idx + 1]] = false;
            }
        }
    }
    return keep;
}
 
using namespace cv::v4d;
 
class PedestrianDemoPlan : public Plan {
    struct Cache {
        cv::UMat blur_;
    } cache_;
    //BGRA
    cv::UMat background_;
    //RGB
    cv::UMat videoFrame_, videoFrameDown_;
    //GREY
    cv::UMat videoFrameDownGrey_;
 
    //detected pedestrian locations rectangles
    std::vector<cv::Rect> locations_;
    //detected pedestrian locations as boxes
    vector<vector<double>> boxes_;
    //probability of detected object being a pedestrian - currently always set to 1.0
    vector<double> probs_;
    //Faster tracking parameters
    cv::TrackerKCF::Params params_;
    //KCF tracker used instead of continous detection
    cv::Ptr<cv::Tracker> tracker_;
    cv::Rect tracked_ = cv::Rect(0,0,1,1);
    bool trackerInitialized_ = false;
    //If tracking fails re-detect
    bool redetect_ = true;
    //Descriptor used for pedestrian detection
    cv::HOGDescriptor hog_;
 
    //post process and add layers together
    static void composite_layers(const cv::UMat background, const cv::UMat foreground, cv::UMat dst, int blurKernelSize, Cache& cache) {
        cv::boxFilter(foreground, cache.blur_, -1, cv::Size(blurKernelSize, blurKernelSize), cv::Point(-1,-1), true, cv::BORDER_REPLICATE);
        cv::add(background, cache.blur_, dst);
    }
public:
    void setup(cv::Ptr<V4D> window) override {
        window->parallel([](cv::TrackerKCF::Params& params, cv::Ptr<cv::Tracker>& tracker, cv::HOGDescriptor& hog){
            params.desc_pca = cv::TrackerKCF::GRAY;
            params.compress_feature = false;
            params.compressed_size = 1;
            tracker = cv::TrackerKCF::create(params);
            hog.setSVMDetector(cv::HOGDescriptor::getDefaultPeopleDetector());
        }, params_, tracker_, hog_);
    }
 
    void infer(cv::Ptr<V4D> window) override {
        static auto always = [](){ return true; };
        static auto doRedect = [](const bool& trackerInit, const bool& redetect){ return !trackerInit || redetect; };
        static auto dontRedect = [](const bool& trackerInit, const bool& redetect){ return trackerInit && !redetect; };
 
        window->branch(always);
        {
            window->capture();
 
            window->fb([](const cv::UMat& frameBuffer, cv::UMat& videoFrame){
                //copy video frame
                cvtColor(frameBuffer,videoFrame,cv::COLOR_BGRA2RGB);
                //downsample video frame for hog_ detection
            }, videoFrame_);
 
            window->parallel([](const cv::UMat& videoFrame, cv::UMat& videoFrameDown, cv::UMat& videoFrameDownGrey, cv::UMat& background){
                cv::resize(videoFrame, videoFrameDown, cv::Size(DOWNSIZE_WIDTH, DOWNSIZE_HEIGHT));
                cv::cvtColor(videoFrameDown, videoFrameDownGrey, cv::COLOR_RGB2GRAY);
                cv::cvtColor(videoFrame, background, cv::COLOR_RGB2BGRA);
            }, videoFrame_, videoFrameDown_, videoFrameDownGrey_, background_);
        }
        window->endbranch(always);
 
        //Try to track the pedestrian (if we currently are tracking one), else re-detect using HOG descriptor
        window->branch(doRedect, trackerInitialized_, redetect_);
        {
            window->parallel([](cv::HOGDescriptor& hog, bool& redetect, cv::UMat& videoFrameDownGrey, std::vector<cv::Rect>& locations, vector<vector<double>>& boxes, vector<double>& probs, cv::Ptr<cv::Tracker>& tracker, cv::Rect& tracked, bool& trackerInitialized){
                redetect = false;
                //Detect pedestrians
                hog.detectMultiScale(videoFrameDownGrey, locations, 0, cv::Size(), cv::Size(), 1.15, 2.0, false);
 
                if (!locations.empty()) {
                    boxes.clear();
                    probs.clear();
                    //collect all found boxes
                    for (const auto &rect : locations) {
                        boxes.push_back( { double(rect.x), double(rect.y), double(rect.x + rect.width), double(rect.y + rect.height) });
                        probs.push_back(1.0);
                    }
 
                    //use nms to filter overlapping boxes (https://medium.com/analytics-vidhya/non-max-suppression-nms-6623e6572536)
                    vector<bool> keep = non_maximal_suppression(&boxes, &probs, 0.1);
                    for (size_t i = 0; i < keep.size(); ++i) {
                        if (keep[i]) {
                            //only track the first pedestrian found
                            tracked = locations[i];
                            break;
                        }
                    }
 
                    if(!trackerInitialized) {
        //              initialize the tracker once
                        tracker->init(videoFrameDownGrey, tracked);
                        trackerInitialized = true;
                    }
                }
            }, hog_, redetect_, videoFrameDownGrey_, locations_, boxes_, probs_, tracker_, tracked_, trackerInitialized_);
        }
        window->endbranch(doRedect, trackerInitialized_, redetect_);
        window->branch(dontRedect, trackerInitialized_, redetect_);
        {
            window->parallel([](bool& redetect, cv::UMat& videoFrameDownGrey, cv::Ptr<cv::Tracker>& tracker, cv::Rect& tracked){
                if(!tracker->update(videoFrameDownGrey, tracked)) {
                    //detection failed - re-detect
                    redetect = true;
                }
            }, redetect_, videoFrameDownGrey_, tracker_, tracked_);
        }
        window->endbranch(dontRedect, trackerInitialized_, redetect_);
 
        window->branch(always);
        {
        //Draw an ellipse around the tracked pedestrian
            window->nvg([](const cv::Size& sz, cv::Rect& tracked) {
                using namespace cv::v4d::nvg;
                clear();
                beginPath();
                strokeWidth(std::fmax(2.0, sz.width / 960.0));
                strokeColor(cv::v4d::colorConvert(cv::Scalar(0, 127, 255, 200), cv::COLOR_HLS2BGR));
                float width = tracked.width * WIDTH_SCALE;
                float height = tracked.height * HEIGHT_SCALE;
                float cx = tracked.x * WIDTH_SCALE + (width / 2.0f);
                float cy = tracked.y * HEIGHT_SCALE + (height / 2.0f);
                ellipse(cx, cy, width / 2.0f, height / 2.0f);
                stroke();
            }, window->fbSize(), tracked_);
 
            //Put it all together
            window->fb([](cv::UMat& frameBuffer, cv::UMat& bg, Cache& cache){
                composite_layers(bg, frameBuffer, frameBuffer, BLUR_KERNEL_SIZE, cache);
            }, background_, cache_);
 
            window->write();
        }
        window->endbranch(always);
    }
};
int main(int argc, char **argv) {
    CV_UNUSED(argc);
    CV_UNUSED(argv);
 
    if (argc != 2) {
        std::cerr << "Usage: pedestrian-demo <video-input>" << endl;
        exit(1);
    }
 
    using namespace cv::v4d;
    cv::Ptr<V4D> window = V4D::make(WIDTH, HEIGHT, "Pedestrian Demo", ALL, OFFSCREEN);
 
    window->printSystemInfo();
 
#ifndef __EMSCRIPTEN__
    auto src = makeCaptureSource(window, argv[1]);
    window->setSource(src);
 
    auto sink = makeWriterSink(window, OUTPUT_FILENAME, src->fps(), cv::Size(WIDTH, HEIGHT));
    window->setSink(sink);
#else
    auto src = makeCaptureSource(WIDTH, HEIGHT, window);
    window->setSource(src);
#endif
 
    window->run<PedestrianDemoPlan>(2);
 
    return 0;
}