417 lines
14 KiB
C++
417 lines
14 KiB
C++
#include "sv_mot.h"
|
|
#include <cmath>
|
|
#include <fstream>
|
|
#include "gason.h"
|
|
#include "sv_util.h"
|
|
|
|
using namespace std;
|
|
using namespace Eigen;
|
|
|
|
namespace sv {
|
|
|
|
|
|
MultipleObjectTracker::MultipleObjectTracker()
|
|
{
|
|
this->_params_loaded = false;
|
|
this->_sort_impl = NULL;
|
|
}
|
|
MultipleObjectTracker::~MultipleObjectTracker()
|
|
{
|
|
if (this->_sort_impl)
|
|
delete this->_sort_impl;
|
|
}
|
|
|
|
void MultipleObjectTracker::track(cv::Mat img_, TargetsInFrame& tgts_)
|
|
{
|
|
if (!this->_params_loaded)
|
|
{
|
|
this->_load();
|
|
this->_params_loaded = true;
|
|
}
|
|
if ("sort" == this->_algorithm && this->_sort_impl)
|
|
{
|
|
this->_detector->detect(img_, tgts_);
|
|
this->_sort_impl->update(tgts_);
|
|
}
|
|
}
|
|
|
|
void MultipleObjectTracker::init(CommonObjectDetector* detector_)
|
|
{
|
|
if (!this->_params_loaded)
|
|
{
|
|
this->_load();
|
|
this->_params_loaded = true;
|
|
}
|
|
if ("sort" == this->_algorithm)
|
|
{
|
|
this->_sort_impl = new SORT(this->_iou_thres, this->_max_age, this->_min_hits);
|
|
}
|
|
this->_detector = detector_;
|
|
}
|
|
|
|
void MultipleObjectTracker::_load()
|
|
{
|
|
JsonValue all_value;
|
|
JsonAllocator allocator;
|
|
_load_all_json(this->alg_params_fn, all_value, allocator);
|
|
|
|
JsonValue tracker_params_value;
|
|
_parser_algorithm_params("MultipleObjectTracker", all_value, tracker_params_value);
|
|
|
|
for (auto i : tracker_params_value) {
|
|
if ("algorithm" == std::string(i->key)) {
|
|
this->_algorithm = i->value.toString();
|
|
std::cout << "algorithm: " << this->_algorithm << std::endl;
|
|
}
|
|
else if ("with_reid" == std::string(i->key)) {
|
|
if (i->value.getTag() == JSON_TRUE)
|
|
this->_with_reid = true;
|
|
else
|
|
this->_with_reid = false;
|
|
|
|
std::cout << "with_reid: " << this->_with_reid << std::endl;
|
|
}
|
|
else if ("reid_input_size" == std::string(i->key) && i->value.getTag() == JSON_ARRAY) {
|
|
int jcnt = 0;
|
|
for (auto j : i->value) {
|
|
if (jcnt == 0) {
|
|
this->_reid_input_w = j->value.toNumber();
|
|
}
|
|
if (jcnt == 1) {
|
|
this->_reid_input_h = j->value.toNumber();
|
|
}
|
|
jcnt += 1;
|
|
}
|
|
std::cout << "reid_input_w: " << this->_reid_input_w << std::endl;
|
|
std::cout << "reid_input_h: " << this->_reid_input_h << std::endl;
|
|
}
|
|
else if ("reid_num_samples" == std::string(i->key)) {
|
|
this->_reid_num_samples = i->value.toNumber();
|
|
std::cout << "reid_num_samples: " << this->_reid_num_samples << std::endl;
|
|
}
|
|
else if ("reid_match_thres" == std::string(i->key)) {
|
|
this->_reid_match_thres = i->value.toNumber();
|
|
std::cout << "reid_match_thres: " << this->_reid_match_thres << std::endl;
|
|
}
|
|
else if ("iou_thres" == std::string(i->key)) {
|
|
this->_iou_thres = i->value.toNumber();
|
|
std::cout << "iou_thres: " << this->_iou_thres << std::endl;
|
|
}
|
|
else if ("max_age" == std::string(i->key)) {
|
|
this->_max_age = i->value.toNumber();
|
|
std::cout << "max_age: " << this->_max_age << std::endl;
|
|
}
|
|
else if ("min_hits" == std::string(i->key)) {
|
|
this->_min_hits = i->value.toNumber();
|
|
std::cout << "min_hits: " << this->_min_hits << std::endl;
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
|
|
KalmanFilter::KalmanFilter()
|
|
{
|
|
this->_chi2inv95 << 3.8415, 5.9915, 7.8147, 9.4877, 11.070, 12.592, 14.067, 15.507, 16.919;
|
|
this->_F = MatrixXd::Identity(8, 8);
|
|
for (int i=0; i<4; i++)
|
|
{
|
|
this->_F(i,i+4) = 1;
|
|
}
|
|
this->_H = MatrixXd::Identity(4, 8);
|
|
this->_std_weight_position = 1. / 20;
|
|
this->_std_weight_vel = 1. / 160;
|
|
}
|
|
|
|
KalmanFilter::~KalmanFilter()
|
|
{
|
|
}
|
|
|
|
pair<Matrix<double, 8, 1>, Matrix<double, 8, 8> > KalmanFilter::initiate(Vector4d &bbox)
|
|
{
|
|
Matrix<double,8,1> mean;
|
|
mean << bbox(0), bbox(1), bbox(2) / bbox(3), bbox(3), 0, 0, 0, 0;
|
|
VectorXd stds(8);
|
|
stds << 2 * this->_std_weight_position * mean(3), 2 * this->_std_weight_position * mean(3), 0.01, 2 * this->_std_weight_position * mean(3), \
|
|
10 * this->_std_weight_vel * mean(3), 10 * this->_std_weight_vel * mean(3), 1e-5, 10 * this->_std_weight_vel * mean(3);
|
|
MatrixXd squared = stds.array().square();
|
|
Matrix<double, 8, 8> covariances;
|
|
covariances = squared.asDiagonal();
|
|
return make_pair(mean, covariances);
|
|
}
|
|
|
|
pair<Matrix<double, 8, 1>, Matrix<double, 8, 8> > KalmanFilter::update(Matrix<double, 8, 1> mean, Matrix<double, 8, 8> covariances, sv::Box &box)
|
|
{
|
|
MatrixXd R;
|
|
Vector4d stds;
|
|
|
|
stds << this->_std_weight_position * mean(3), this->_std_weight_position * mean(3), 0.1, this->_std_weight_position * mean(3);
|
|
MatrixXd squared = stds.array().square();
|
|
R = squared.asDiagonal();
|
|
MatrixXd S = this->_H * covariances * this->_H.transpose() + R;
|
|
|
|
MatrixXd Kalman_gain = covariances * this->_H.transpose() * S.inverse();
|
|
VectorXd measurement(4);
|
|
measurement << box.x1, box.y1, (box.x2-box.x1) / (box.y2-box.y1), box.y2-box.y1;
|
|
Matrix<double, 8, 1> new_mean = mean + Kalman_gain * (measurement - this->_H * mean);
|
|
Matrix<double, 8, 8> new_covariances = (MatrixXd::Identity(8, 8) - Kalman_gain * this->_H) * covariances;
|
|
return make_pair(new_mean, new_covariances);
|
|
}
|
|
|
|
pair<Matrix<double, 8, 1>, Matrix<double, 8, 8> > KalmanFilter::predict(Matrix<double, 8, 1> mean, Matrix<double, 8, 8> covariances)
|
|
{
|
|
VectorXd stds(8);
|
|
stds << this->_std_weight_position * mean(3), this->_std_weight_position * mean(3), 0.01, this->_std_weight_position * mean(3), \
|
|
this->_std_weight_vel * mean(3), this->_std_weight_vel * mean(3), 1e-5, this->_std_weight_vel * mean(3);
|
|
MatrixXd squared = stds.array().square();
|
|
MatrixXd Q = squared.asDiagonal();
|
|
Matrix<double, 8, 1> pre_mean = this->_F * mean;
|
|
Matrix<double, 8, 8> pre_cov = this->_F * covariances * this->_F.transpose() + Q;
|
|
return make_pair(pre_mean, pre_cov);
|
|
}
|
|
|
|
|
|
SORT::~SORT()
|
|
{
|
|
}
|
|
|
|
void SORT::update(TargetsInFrame& tgts)
|
|
{
|
|
sv::KalmanFilter kf;
|
|
if (! this->_tracklets.size())
|
|
{
|
|
Vector4d bbox;
|
|
for (int i=0; i<tgts.targets.size(); i++)
|
|
{
|
|
sv::Box box;
|
|
tgts.targets[i].getBox(box);
|
|
Tracklet tracklet;
|
|
tracklet.id = ++ this->_next_tracklet_id;
|
|
// cout << tracklet.id << endl;
|
|
tgts.targets[i].tracked_id = this->_next_tracklet_id;
|
|
tgts.targets[i].has_tid = true;
|
|
|
|
tracklet.bbox << box.x1, box.y1, box.x2-box.x1, box.y2-box.y1; // x,y,w,h
|
|
tracklet.age = 0;
|
|
tracklet.hits = 1;
|
|
tracklet.misses = 0;
|
|
// initate the motion
|
|
pair<Matrix<double, 8, 1>, Matrix<double, 8, 8> > motion = kf.initiate(tracklet.bbox);
|
|
tracklet.mean = motion.first;
|
|
tracklet.covariance = motion.second;
|
|
|
|
this->_tracklets.push_back(tracklet);
|
|
}
|
|
}
|
|
else
|
|
{
|
|
for (int i=0; i<tgts.targets.size(); i++)
|
|
{
|
|
tgts.targets[i].tracked_id = 0;
|
|
tgts.targets[i].has_tid = true;
|
|
}
|
|
|
|
array<int, 100> match_det;
|
|
match_det.fill(-1);
|
|
// predict the next state of each tracklet
|
|
for (auto& tracklet : this->_tracklets)
|
|
{
|
|
tracklet.age++;
|
|
pair<Matrix<double, 8, 1>, Matrix<double, 8, 8> > motion = kf.predict(tracklet.mean, tracklet.covariance);
|
|
tracklet.bbox << motion.first(0), motion.first(1), motion.first(2) * motion.first(3), motion.first(3);
|
|
tracklet.mean = motion.first;
|
|
tracklet.covariance = motion.second;
|
|
}
|
|
|
|
// Match the detections to the existing tracklets
|
|
// cout << "the num of targets: " << tgts.targets.size() << endl;
|
|
// cout << "the num of tracklets: " << this->_tracklets.size() << endl;
|
|
vector<vector<double> > iouMatrix(this->_tracklets.size(), vector<double> (tgts.targets.size(), 0));
|
|
for (int i=0; i<this->_tracklets.size(); i++)
|
|
{
|
|
for (int j=0; j<tgts.targets.size(); j++)
|
|
{
|
|
sv::Box box;
|
|
tgts.targets[j].getBox(box);
|
|
iouMatrix[i][j] = this->_iou(this->_tracklets[i], box);
|
|
}
|
|
}
|
|
vector<pair<int, int> > matches = this->_hungarian(iouMatrix);
|
|
for (auto& match : matches)
|
|
{
|
|
int trackletIndex = match.first;
|
|
int detectionIndex = match.second;
|
|
if (trackletIndex >= 0 && detectionIndex >= 0)
|
|
{
|
|
if (iouMatrix[match.first][match.second] >= 0)
|
|
{
|
|
sv::Box box;
|
|
tgts.targets[detectionIndex].getBox(box);
|
|
this->_tracklets[trackletIndex].age = 0;
|
|
this->_tracklets[trackletIndex].hits++;
|
|
this->_tracklets[trackletIndex].bbox << box.x1, box.y1, box.x2-box.x1, box.y2-box.y1;
|
|
|
|
auto[mean, covariance] = kf.update(this->_tracklets[trackletIndex].mean, this->_tracklets[trackletIndex].covariance, box);
|
|
this->_tracklets[trackletIndex].mean = mean;
|
|
this->_tracklets[trackletIndex].covariance = covariance;
|
|
|
|
tgts.targets[detectionIndex].tracked_id = this->_tracklets[trackletIndex].id;
|
|
match_det[detectionIndex] = detectionIndex;
|
|
}
|
|
}
|
|
}
|
|
// create new tracklets for unmatched detections
|
|
for (int i = 0; i < tgts.targets.size(); i++)
|
|
{
|
|
if (match_det[i] == -1)
|
|
{
|
|
sv::Box box;
|
|
tgts.targets[i].getBox(box);
|
|
Tracklet tracklet;
|
|
tracklet.id = ++ this->_next_tracklet_id;
|
|
tracklet.bbox << box.x1, box.y1, box.x2-box.x1, box.y2-box.y1;
|
|
|
|
tracklet.age = 0;
|
|
tracklet.hits = 1;
|
|
tracklet.misses = 0;
|
|
|
|
auto[new_mean, new_covariance] = kf.initiate(tracklet.bbox);
|
|
tracklet.mean = new_mean;
|
|
tracklet.covariance = new_covariance;
|
|
|
|
tgts.targets[i].tracked_id = this->_next_tracklet_id;
|
|
tgts.targets[i].has_tid = true;
|
|
this->_tracklets.push_back(tracklet);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
vector<Tracklet> SORT::getTracklets() const
|
|
{
|
|
return this->_tracklets;
|
|
}
|
|
|
|
double SORT::_iou(Tracklet& tracklet, sv::Box& box)
|
|
{
|
|
double trackletX1 = tracklet.bbox(0);
|
|
double trackletY1 = tracklet.bbox(1);
|
|
double trackletX2 = tracklet.bbox(0) + tracklet.bbox(2);
|
|
double trackletY2 = tracklet.bbox(1) + tracklet.bbox(3);
|
|
|
|
double detectionX1 = box.x1;
|
|
double detectionY1 = box.y1;
|
|
double detectionX2 = box.x2;
|
|
double detectionY2 = box.y2;
|
|
double intersectionX1 = max(trackletX1, detectionX1);
|
|
double intersectionY1 = max(trackletY1, detectionY1);
|
|
double intersectionX2 = min(trackletX2, detectionX2);
|
|
double intersectionY2 = min(trackletY2, detectionY2);
|
|
|
|
double w = (intersectionX2-intersectionX1 > 0.0) ? (intersectionX2-intersectionX1) : 0.0;
|
|
double h = (intersectionY2-intersectionY1 > 0.0) ? (intersectionY2-intersectionY1) : 0.0;
|
|
double intersectionArea = w * h;
|
|
|
|
double trackletArea = tracklet.bbox(2) * tracklet.bbox(3);
|
|
|
|
double detectionArea = (box.x2-box.x1) * (box.y2-box.y1);
|
|
double unionArea = trackletArea + detectionArea - intersectionArea;
|
|
double iou = (1 - intersectionArea / unionArea * 1.0);
|
|
|
|
return iou;
|
|
}
|
|
|
|
vector<pair<int, int> > SORT::_hungarian(vector<vector<double> > costMatrix)
|
|
{
|
|
int numRows = costMatrix.size();
|
|
int numCols = costMatrix[0].size();
|
|
|
|
const bool transposed = numCols > numRows;
|
|
// transpose the matrix if necessary
|
|
if (transposed)
|
|
{
|
|
vector<vector<double> > transposedMatrix(numCols, vector<double>(numRows));
|
|
for (int i=0; i<numRows; i++)
|
|
{
|
|
for (int j=0; j<numCols; j++)
|
|
{
|
|
transposedMatrix[j][i] = costMatrix[i][j];
|
|
}
|
|
}
|
|
costMatrix = transposedMatrix;
|
|
swap(numRows, numCols);
|
|
}
|
|
vector<double>rowMin(numRows, numeric_limits<double>::infinity());
|
|
vector<double>colMin(numCols, numeric_limits<double>::infinity());
|
|
vector<int>rowMatch(numRows, -1);
|
|
vector<int>colMatch(numCols, -1);
|
|
vector<pair<int, int> > matches;
|
|
// step1: Subtract the row minimums from each row
|
|
for (int i=0; i<numRows; i++)
|
|
{
|
|
for (int j=0; j<numCols; j++)
|
|
{
|
|
rowMin[i] = min(rowMin[i], costMatrix[i][j]);
|
|
}
|
|
for (int j=0; j<numCols; j++)
|
|
{
|
|
costMatrix[i][j] -= rowMin[i];
|
|
}
|
|
}
|
|
// step2: substract the colcum minimums from each column
|
|
for (int j=0; j<numCols; j++)
|
|
{
|
|
for (int i=0; i<numRows; i++)
|
|
{
|
|
colMin[j] = min(colMin[j], costMatrix[i][j]);
|
|
}
|
|
for (int i=0; i<numRows; i++)
|
|
{
|
|
costMatrix[i][j] -= colMin[j];
|
|
}
|
|
}
|
|
// step3: find a maximal matching
|
|
for (int i=0; i<numRows; i++)
|
|
{
|
|
vector<bool> visited(numCols, false);
|
|
this->_augment(costMatrix, i, rowMatch, colMatch, visited);
|
|
}
|
|
// step4: calculate the matches
|
|
matches.clear();
|
|
for (int j=0; j<numCols; j++)
|
|
{
|
|
matches.push_back(make_pair(colMatch[j], j));
|
|
}
|
|
if (transposed)
|
|
{
|
|
for (auto& match : matches)
|
|
{
|
|
swap(match.first, match.second);
|
|
}
|
|
}
|
|
return matches;
|
|
}
|
|
|
|
bool SORT::_augment(const vector<vector<double> >& costMatrix, int row, vector<int>& rowMatch, vector<int>& colMatch, vector<bool>& visited)
|
|
{
|
|
int numCols = costMatrix[0].size();
|
|
for (int j=0; j<numCols; j++)
|
|
{
|
|
if (costMatrix[row][j] == 0 && !visited[j])
|
|
{
|
|
visited[j] = true;
|
|
if (colMatch[j] == -1 || this->_augment(costMatrix, colMatch[j], rowMatch, colMatch, visited))
|
|
{
|
|
rowMatch[row] = j;
|
|
colMatch[j] = row;
|
|
return true;
|
|
}
|
|
}
|
|
}
|
|
return false;
|
|
}
|
|
|
|
|
|
}
|
|
|