1 #include "MultiTracker.h"
7 using namespace suanzi;
11 static const std::string TAG = "MultiTracker";
12 static const cv::Size PREFERRED_SIZE = Size(64, 128);
14 #define MaxCost 100000
16 MultiTracker::MultiTracker(EngineWPtr e)
19 LOG_DEBUG(TAG, "init - loading model.pkl");
20 predictor = PredictorWrapper::create("./python/model.pkl");
22 this->descriptor = {Size(64, 128), Size(16, 16), Size(8, 8), Size(8, 8), 9};
24 std::vector<double> ff (40, 1);
25 double prob = predictor->predict(4, ff);
28 MultiTracker::~MultiTracker()
34 static double calc_iou_ratio(const Detection& d1, const Detection& d2)
40 static std::vector<double> similarity(const PatchPtr p1, const PatchPtr p2)
42 std::vector<double> feature;
43 cv::Mat im1(PREFERRED_SIZE, p1->image_crop.type());
44 cv::Mat im2(PREFERRED_SIZE, p2->image_crop.type());
45 cv::resize(p1->image_crop, im1, im1.size());
46 cv::resize(p2->image_crop, im2, im2.size());
48 cv::matchTemplate(im1, im2, result, CV_TM_CCOEFF_NORMED);
49 feature.push_back(result.at<double>(0, 0));
50 cv::matchTemplate(im1, im2, result, CV_TM_CCORR_NORMED);
51 feature.push_back(result.at<double>(0, 0));
54 vector<double>& f1_hog = p1->features.first; Mat f1_hue = p1->features.second;
55 vector<double>& f2_hog = p1->features.first; Mat f2_hue = p1->features.second;
56 feature.push_back(distance_cosine(Eigen::Map<Eigen::VectorXd>(f1_hog.data(), f1_hog.size()),
57 Eigen::Map<Eigen::VectorXd>(f2_hog.data(), f2_hog.size())));
58 feature.push_back(distance_euclidean(Eigen::Map<Eigen::VectorXd>(f1_hog.data(), f1_hog.size()),
59 Eigen::Map<Eigen::VectorXd>(f2_hog.data(), f2_hog.size())));
60 feature.push_back(compareHist(f1_hue, f2_hue, HISTCMP_CORREL));
61 feature.push_back(compareHist(f1_hue, f2_hue, HISTCMP_HELLINGER));
63 Detection& d1 = p1->detection;
64 Detection& d2 = p2->detection;
66 double center_distance = sqrt(pow((d1.center_x - d2.center_x), 2) + pow((d1.center_y - d2.center_y), 2));
67 feature.push_back(center_distance / (d1.width + d1.height + d2.width + d2.height) * 4);
69 double iou_ratio = calc_iou_ratio(d1, d2);
70 feature.push_back(iou_ratio);
76 double MultiTracker::distance(TrackerPtr tracker, const cv::Mat& image, const Detection& d)
78 PatchPtr patch = createPatch(image, d);
79 std::vector<double> features;
81 std::vector<double> ss;
82 for (auto i : tracker->patches){
83 ss = similarity(i, patch);
84 features.insert(features.end(), ss.begin(), ss.end());
86 double prob = predictor->predict(4, features);
92 void MultiTracker::update(unsigned int total, const Detection* detections, const Mat& image)
96 if (EnginePtr e = engine.lock()){
103 int row = trackers.size();
105 Eigen::MatrixXi cost_matrix = Eigen::MatrixXi::Zero(row, col);
106 for (int i = 0; i < row; i++){
107 for (int j = 0; j < col; j++){
108 //if (calc_iou_ratio(trackers[i], detections[j]) < -0.1)
109 // cost_matrix(i, j) = MaxCost;
111 cost_matrix(i, j) = distance(trackers[i], image, detections[j]);
115 Eigen::VectorXi tracker_inds, bb_inds;
116 linear_sum_assignment(cost_matrix, tracker_inds, bb_inds);
118 // handle unmatched trackers
119 //vector<TrackerPtr> unmatched_trackers;
120 for (int i = 0; i < row; i++){
121 if (!(tracker_inds.array() == i).any()){
122 trackers[i]->updateState(image);
126 // handle unmatched detections
127 vector<int> unmatched_detection;
128 for(int j = 0; j < col; j++){
129 if (!(bb_inds.array() == j).any()){
130 unmatched_detection.push_back(j);
133 // create new trackers for new detections
134 for (auto i : unmatched_detection){
135 TrackerPtr t (new Tracker(image));
136 this->trackers.push_back(t);
141 PatchPtr pp = createPatch(image, dd);
144 // Get image crop from input image within given bounding box - Detecinon
145 static cv::Mat image_crop(const cv::Mat& image, const Detection& bb)
149 return image.clone();
152 PatchPtr MultiTracker::createPatch(const Mat& image, const Detection& detect)
154 PatchPtr patch(new Patch());
156 // calculate hog descriptors, size is 3780
158 im = image_crop(image, detect);
159 resize(im, im2, PREFERRED_SIZE);
160 vector<float> feature_hog;
161 this->descriptor.compute(im2, feature_hog);
163 // calculate histogram, size is (64 x 45)
165 cvtColor(im, hsv, COLOR_BGR2HSV);
166 int channels[] = {0, 1};
167 int histSize[] = {45, 64};
168 float hranges[] = {0, 180};
169 float sranges[] = {0, 256};
170 const float* ranges[] = {hranges, sranges};
171 calcHist(&hsv, 1, channels, Mat(), hist, 2, histSize, ranges, true, false);
173 patch->image_crop = im.clone();
174 patch->detection = detect;
175 std::vector<double> feature_hog_double (feature_hog.begin(), feature_hog.end()); // convert to double
176 patch->features = std::make_pair(feature_hog_double, hist);