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 = PredictorWrapper::create("./python", "./python/model.pkl");
23 this->descriptor = {Size(64, 128), Size(16, 16), Size(8, 8), Size(8, 8), 9};
26 MultiTracker::~MultiTracker()
32 static std::vector<double> similarity(const PatchPtr p1, const PatchPtr p2)
34 std::vector<double> feature;
35 cv::Mat im1(PREFERRED_SIZE, p1->image_crop.type());
36 cv::Mat im2(PREFERRED_SIZE, p2->image_crop.type());
37 cv::resize(p1->image_crop, im1, im1.size());
38 cv::resize(p2->image_crop, im2, im2.size());
40 cv::matchTemplate(im1, im2, result, CV_TM_CCOEFF_NORMED);
41 feature.push_back(result.at<double>(0, 0));
42 cv::matchTemplate(im1, im2, result, CV_TM_CCORR_NORMED);
43 feature.push_back(result.at<double>(0, 0));
46 vector<double>& f1_hog = p1->features.first; Mat f1_hue = p1->features.second;
47 vector<double>& f2_hog = p1->features.first; Mat f2_hue = p1->features.second;
48 feature.push_back(distance_cosine(Eigen::Map<Eigen::VectorXd>(f1_hog.data(), f1_hog.size()),
49 Eigen::Map<Eigen::VectorXd>(f2_hog.data(), f2_hog.size())));
50 feature.push_back(distance_euclidean(Eigen::Map<Eigen::VectorXd>(f1_hog.data(), f1_hog.size()),
51 Eigen::Map<Eigen::VectorXd>(f2_hog.data(), f2_hog.size())));
52 feature.push_back(compareHist(f1_hue, f2_hue, HISTCMP_CORREL));
53 feature.push_back(compareHist(f1_hue, f2_hue, HISTCMP_HELLINGER));
55 Detection& d1 = p1->detection;
56 Detection& d2 = p2->detection;
58 double center_distance = sqrt(pow((d1.center_x - d2.center_x), 2) + pow((d1.center_y - d2.center_y), 2));
59 feature.push_back(center_distance / (d1.width + d1.height + d2.width + d2.height) * 4);
61 feature.push_back(calc_iou_ratio(getRectInDetection(d1), getRectInDetection(d2)));
66 double MultiTracker::distance(TrackerPtr tracker, const cv::Mat& image, const Detection& d)
68 PatchPtr patch = createPatch(image, d);
69 std::vector<double> features;
71 std::vector<double> ss;
72 for (auto i : tracker->patches){
73 ss = similarity(i, patch);
74 features.insert(features.end(), ss.begin(), ss.end());
76 double prob = predictor->predict(Tracker::MaxPatch - 1, features); // TODO why is MaxPatch-1
77 if (prob > ProbThreshold)
83 static float calc_iou_ratio(const Detection& d1, const Detection& d2)
85 return calc_iou_ratio(getRectInDetection(d1), getRectInDetection(d2));
88 void MultiTracker::update(unsigned int total, const Detection* detections, const Mat& image)
90 // predict trackers, update trackers using kalman filter
91 for (auto t : trackers){
95 // match the trackers with the detections using linear sum assignment (hungarian)
96 int row = trackers.size();
98 Eigen::MatrixXi cost_matrix = Eigen::MatrixXi::Zero(row, col);
99 for (int i = 0; i < row; i++){
100 for (int j = 0; j < col; j++){
101 if (calc_iou_ratio(trackers[i]->detection, detections[j]) < -0.1)
102 cost_matrix(i, j) = MaxCost;
104 cost_matrix(i, j) = distance(trackers[i], image, detections[j]);
108 Eigen::VectorXi tracker_inds, bb_inds;
109 linear_sum_assignment(cost_matrix, tracker_inds, bb_inds);
111 set<TrackerPtr> unmatched_trackers;
112 set<int> unmatch_bbs_indices;
114 for(unsigned int i = 0; i < trackers.size(); i++){
115 if (!(tracker_inds.array() == i).any()){
116 unmatched_trackers.insert(trackers[i]);
119 for (unsigned int j = 0; j < total; j++){
120 if (!(bb_inds.array() == j).any()){
121 unmatch_bbs_indices.insert(j);
125 // handle matched trackers
126 for (unsigned int i = 0; i < tracker_inds.size(); i++){
127 for (int j = 0; j < bb_inds.size(); j++){
128 int rr = tracker_inds(i);
130 TrackerPtr tracker = trackers[rr];
131 const Detection& detect = detections[cc];
132 if (cost_matrix(rr, cc) < MaxCost){
133 tracker->correct(image, detect);
134 tracker->addPatch(createPatch(image, detect));
136 unmatched_trackers.insert(tracker); // failed trackers
137 unmatch_bbs_indices.insert(cc); // filed detection
142 // handle unmatched trackers
143 for (auto t : unmatched_trackers){
144 t->updateState(image);
147 // handle unmatched detections - Create new trackers
148 vector<Person> inPersons;
149 for (auto i : unmatch_bbs_indices){
150 TrackerPtr new_tracker (new Tracker(image, detections[i]));
151 new_tracker->addPatch(createPatch(image, detections[i]));
152 this->trackers.push_back(new_tracker);
154 inPersons.push_back(test);
157 // callback and notify engine - persons in
158 if (inPersons.size() > 0){
159 if (auto e = engine.lock()){
160 e->onPersonsIn(inPersons);
164 // Delete lost trackers
165 vector<Person> outPersons;
166 for (auto it = trackers.begin(); it < trackers.end(); it++){
167 if ((*it)->status == TrackerStatus::Delete){
169 outPersons.push_back(test);
174 // callback and notify engine - persons out
175 if (outPersons.size() > 0){
176 if (auto e = engine.lock()){
177 e->onPersonsOut(outPersons);
182 static cv::Mat image_crop(const cv::Mat& image, const Detection& bb)
184 return image(getRectInDetection(bb));
187 PatchPtr MultiTracker::createPatch(const Mat& image, const Detection& detect)
189 PatchPtr patch(new Patch());
191 // calculate hog descriptors, size is 3780
193 im = image_crop(image, detect);
194 resize(im, im2, PREFERRED_SIZE);
195 vector<float> feature_hog;
196 this->descriptor.compute(im2, feature_hog);
198 // calculate histogram, size is (64 x 45)
200 cvtColor(im, hsv, COLOR_BGR2HSV);
201 int channels[] = {0, 1};
202 int histSize[] = {45, 64};
203 float hranges[] = {0, 180};
204 float sranges[] = {0, 256};
205 const float* ranges[] = {hranges, sranges};
206 calcHist(&hsv, 1, channels, Mat(), hist, 2, histSize, ranges, true, false);
208 patch->image_crop = im.clone();
209 patch->detection = detect;
210 std::vector<double> feature_hog_double (feature_hog.begin(), feature_hog.end()); // convert to double
211 patch->features = std::make_pair(feature_hog_double, hist);