1 #include "MultiTracker.h"
8 using namespace suanzi;
12 static const std::string TAG = "MultiTracker";
13 static const cv::Size PREFERRED_SIZE = Size(64, 128);
14 static const double MaxCost = 100000;
15 static const double ProbThreshold = 0.05;
17 MultiTracker::MultiTracker(EngineWPtr e)
20 LOG_DEBUG(TAG, "init - loading model.pkl");
21 predictor = PredictorWrapperPtr(new PredictorWrapper());
22 predictor->load("./resources/model.pkl");
24 this->descriptor = {Size(64, 128), Size(16, 16), Size(8, 8), Size(8, 8), 9};
27 MultiTracker::~MultiTracker()
33 static std::vector<double> similarity(const PatchPtr p1, const PatchPtr p2)
35 std::vector<double> feature;
36 cv::Mat im1(PREFERRED_SIZE, p1->image_crop.type());
37 cv::Mat im2(PREFERRED_SIZE, p2->image_crop.type());
38 cv::resize(p1->image_crop, im1, im1.size());
39 cv::resize(p2->image_crop, im2, im2.size());
41 cv::matchTemplate(im1, im2, result, CV_TM_CCOEFF_NORMED);
42 feature.push_back(result.at<double>(0, 0));
43 cv::matchTemplate(im1, im2, result, CV_TM_CCORR_NORMED);
44 feature.push_back(result.at<double>(0, 0));
47 vector<double>& f1_hog = p1->features.first; Mat f1_hue = p1->features.second;
48 vector<double>& f2_hog = p1->features.first; Mat f2_hue = p1->features.second;
49 feature.push_back(distance_cosine(Eigen::Map<Eigen::VectorXd>(f1_hog.data(), f1_hog.size()),
50 Eigen::Map<Eigen::VectorXd>(f2_hog.data(), f2_hog.size())));
51 feature.push_back(distance_euclidean(Eigen::Map<Eigen::VectorXd>(f1_hog.data(), f1_hog.size()),
52 Eigen::Map<Eigen::VectorXd>(f2_hog.data(), f2_hog.size())));
53 feature.push_back(compareHist(f1_hue, f2_hue, HISTCMP_CORREL));
54 feature.push_back(compareHist(f1_hue, f2_hue, HISTCMP_HELLINGER));
56 Detection& d1 = p1->detection;
57 Detection& d2 = p2->detection;
59 double center_distance = sqrt(pow((d1.center_x - d2.center_x), 2) + pow((d1.center_y - d2.center_y), 2));
60 feature.push_back(center_distance / (d1.width + d1.height + d2.width + d2.height) * 4);
62 feature.push_back(calc_iou_ratio(getRectInDetection(d1), getRectInDetection(d2)));
67 double MultiTracker::distance(TrackerPtr tracker, const cv::Mat& image, const Detection& d)
69 PatchPtr patch = createPatch(image, d);
70 std::vector<double> features;
72 std::vector<double> ss;
73 for (auto i : tracker->patches){
74 ss = similarity(i, patch);
75 features.insert(features.end(), ss.begin(), ss.end());
77 double prob = predictor->predict(Tracker::MaxPatch - 1, features); // TODO why is MaxPatch-1
78 if (prob > ProbThreshold)
84 static float calc_iou_ratio(const Detection& d1, const Detection& d2)
86 return calc_iou_ratio(getRectInDetection(d1), getRectInDetection(d2));
89 void MultiTracker::update(unsigned int total, const Detection* detections, const Mat& image)
91 // predict trackers, update trackers using kalman filter
92 for (auto t : trackers){
96 // match the trackers with the detections using linear sum assignment (hungarian)
97 int row = trackers.size();
99 Eigen::MatrixXi cost_matrix = Eigen::MatrixXi::Zero(row, col);
100 for (int i = 0; i < row; i++){
101 for (int j = 0; j < col; j++){
102 if (calc_iou_ratio(trackers[i]->detection, detections[j]) < -0.1)
103 cost_matrix(i, j) = MaxCost;
105 cost_matrix(i, j) = distance(trackers[i], image, detections[j]);
109 Eigen::VectorXi tracker_inds, bb_inds;
110 linear_sum_assignment(cost_matrix, tracker_inds, bb_inds);
112 set<TrackerPtr> unmatched_trackers;
113 set<int> unmatch_bbs_indices;
115 for(unsigned int i = 0; i < trackers.size(); i++){
116 if (!(tracker_inds.array() == i).any()){
117 unmatched_trackers.insert(trackers[i]);
120 for (unsigned int j = 0; j < total; j++){
121 if (!(bb_inds.array() == j).any()){
122 unmatch_bbs_indices.insert(j);
126 // handle matched trackers
127 for (unsigned int i = 0; i < tracker_inds.size(); i++){
128 for (int j = 0; j < bb_inds.size(); j++){
129 int rr = tracker_inds(i);
131 TrackerPtr tracker = trackers[rr];
132 const Detection& detect = detections[cc];
133 if (cost_matrix(rr, cc) < MaxCost){
134 tracker->correct(image, detect);
135 tracker->addPatch(createPatch(image, detect));
137 unmatched_trackers.insert(tracker); // failed trackers
138 unmatch_bbs_indices.insert(cc); // filed detection
143 // handle unmatched trackers
144 for (auto t : unmatched_trackers){
145 t->updateState(image);
148 // handle unmatched detections - Create new trackers
149 vector<Person> inPersons;
150 for (auto i : unmatch_bbs_indices){
151 TrackerPtr new_tracker (new Tracker(image, detections[i]));
152 new_tracker->addPatch(createPatch(image, detections[i]));
153 this->trackers.push_back(new_tracker);
155 inPersons.push_back(test);
158 // callback and notify engine - persons in
159 if (inPersons.size() > 0){
160 if (auto e = engine.lock()){
161 e->onPersonsIn(inPersons);
165 // Delete lost trackers
166 vector<Person> outPersons;
167 for (auto it = trackers.begin(); it < trackers.end(); it++){
168 if ((*it)->status == TrackerStatus::Delete){
170 outPersons.push_back(test);
175 // callback and notify engine - persons out
176 if (outPersons.size() > 0){
177 if (auto e = engine.lock()){
178 e->onPersonsOut(outPersons);
183 static cv::Mat image_crop(const cv::Mat& image, const Detection& bb)
185 return image(getRectInDetection(bb));
188 PatchPtr MultiTracker::createPatch(const Mat& image, const Detection& detect)
190 PatchPtr patch(new Patch());
192 // calculate hog descriptors, size is 3780
194 im = image_crop(image, detect);
195 resize(im, im2, PREFERRED_SIZE);
196 vector<float> feature_hog;
197 this->descriptor.compute(im2, feature_hog);
199 // calculate histogram, size is (64 x 45)
201 cvtColor(im, hsv, COLOR_BGR2HSV);
202 int channels[] = {0, 1};
203 int histSize[] = {45, 64};
204 float hranges[] = {0, 180};
205 float sranges[] = {0, 256};
206 const float* ranges[] = {hranges, sranges};
207 calcHist(&hsv, 1, channels, Mat(), hist, 2, histSize, ranges, true, false);
209 patch->image_crop = im.clone();
210 patch->detection = detect;
211 std::vector<double> feature_hog_double (feature_hog.begin(), feature_hog.end()); // convert to double
212 patch->features = std::make_pair(feature_hog_double, hist);