costmap_to_dynamic_obstacles.cpp

Go to the documentation of this file.

#include <costmap_converter/costmap_to_dynamic_obstacles/costmap_to_dynamic_obstacles.h>

#include <pluginlib/class_list_macros.h>
#include <tf/tf.h>

PLUGINLIB_EXPORT_CLASS(costmap_converter::CostmapToDynamicObstacles, costmap_converter::BaseCostmapToPolygons)

namespace costmap_converter
{

CostmapToDynamicObstacles::CostmapToDynamicObstacles() : BaseCostmapToPolygons()
{
  ego_vel_.x = ego_vel_.y = ego_vel_.z = 0;
  costmap_ = nullptr;
  dynamic_recfg_ = nullptr;
}

CostmapToDynamicObstacles::~CostmapToDynamicObstacles()
{
  if(dynamic_recfg_ != nullptr)
    delete dynamic_recfg_;
}

void CostmapToDynamicObstacles::initialize(ros::NodeHandle nh)
{
  costmap_ = nullptr;

  // We need the odometry from the robot to compensate the ego motion
  ros::NodeHandle gn; // create odom topic w.r.t. global node handle
  odom_sub_ = gn.subscribe(odom_topic_, 1, &CostmapToDynamicObstacles::odomCallback, this);

  nh.param("publish_static_obstacles", publish_static_obstacles_, publish_static_obstacles_);

  // Foreground detection parameters
  BackgroundSubtractor::Params bg_sub_params;

  bg_sub_params.alpha_slow = 0.3;
  nh.param("alpha_slow", bg_sub_params.alpha_slow, bg_sub_params.alpha_slow);

  bg_sub_params.alpha_fast = 0.85;
  nh.param("alpha_fast", bg_sub_params.alpha_fast, bg_sub_params.alpha_fast);

  bg_sub_params.beta = 0.85;
  nh.param("beta", bg_sub_params.beta, bg_sub_params.beta);

  bg_sub_params.min_occupancy_probability = 180;
  nh.param("min_occupancy_probability", bg_sub_params.min_occupancy_probability, bg_sub_params.min_occupancy_probability);

  bg_sub_params.min_sep_between_fast_and_slow_filter = 80;
  nh.param("min_sep_between_slow_and_fast_filter", bg_sub_params.min_sep_between_fast_and_slow_filter, bg_sub_params.min_sep_between_fast_and_slow_filter);

  bg_sub_params.max_occupancy_neighbors = 100;
  nh.param("max_occupancy_neighbors", bg_sub_params.max_occupancy_neighbors, bg_sub_params.max_occupancy_neighbors);

  bg_sub_params.morph_size = 1;
  nh.param("morph_size", bg_sub_params.morph_size, bg_sub_params.morph_size);

  bg_sub_ = std::unique_ptr<BackgroundSubtractor>(new BackgroundSubtractor(bg_sub_params));

  // Blob detection parameters
  BlobDetector::Params blob_det_params;

  blob_det_params.filterByColor = true; // actually filterByIntensity, always true
  blob_det_params.blobColor = 255;      // Extract light blobs
  blob_det_params.thresholdStep = 256;  // Input for blob detection is already a binary image
  blob_det_params.minThreshold = 127;
  blob_det_params.maxThreshold = 255;
  blob_det_params.minRepeatability = 1;

  blob_det_params.minDistBetweenBlobs = 10;
  nh.param("min_distance_between_blobs", blob_det_params.minDistBetweenBlobs, blob_det_params.minDistBetweenBlobs);

  blob_det_params.filterByArea = true;
  nh.param("filter_by_area", blob_det_params.filterByArea, blob_det_params.filterByArea);

  blob_det_params.minArea = 3; // Filter out blobs with less pixels
  nh.param("min_area", blob_det_params.minArea, blob_det_params.minArea);

  blob_det_params.maxArea = 300;
  nh.param("max_area", blob_det_params.maxArea, blob_det_params.maxArea);

  blob_det_params.filterByCircularity = true; // circularity = 4*pi*area/perimeter^2
  nh.param("filter_by_circularity", blob_det_params.filterByCircularity, blob_det_params.filterByCircularity);

  blob_det_params.minCircularity = 0.2;
  nh.param("min_circularity", blob_det_params.minCircularity, blob_det_params.minCircularity);

  blob_det_params.maxCircularity = 1; // maximal 1 (in case of a circle)
  nh.param("max_circularity", blob_det_params.maxCircularity, blob_det_params.maxCircularity);

  blob_det_params.filterByInertia = true; // Filter blobs based on their elongation
  nh.param("filter_by_intertia", blob_det_params.filterByInertia, blob_det_params.filterByInertia);

  blob_det_params.minInertiaRatio = 0.2;  // minimal 0 (in case of a line)
  nh.param("min_inertia_ratio", blob_det_params.minInertiaRatio, blob_det_params.minInertiaRatio);

  blob_det_params.maxInertiaRatio = 1;    // maximal 1 (in case of a circle)
  nh.param("max_intertia_ratio", blob_det_params.maxInertiaRatio, blob_det_params.maxInertiaRatio);

  blob_det_params.filterByConvexity = false; // Area of the Blob / Area of its convex hull
  nh.param("filter_by_convexity", blob_det_params.filterByConvexity, blob_det_params.filterByConvexity);

  blob_det_params.minConvexity = 0;          // minimal 0
  nh.param("min_convexity", blob_det_params.minConvexity, blob_det_params.minConvexity);

  blob_det_params.maxConvexity = 1;          // maximal 1
  nh.param("max_convexity", blob_det_params.maxConvexity, blob_det_params.maxConvexity);

  blob_det_ = BlobDetector::create(blob_det_params);

  // Tracking parameters
  CTracker::Params tracker_params;
  tracker_params.dt = 0.2;
  nh.param("dt", tracker_params.dt, tracker_params.dt);

  tracker_params.dist_thresh = 60.0;
  nh.param("dist_thresh", tracker_params.dist_thresh, tracker_params.dist_thresh);

  tracker_params.max_allowed_skipped_frames = 3;
  nh.param("max_allowed_skipped_frames", tracker_params.max_allowed_skipped_frames, tracker_params.max_allowed_skipped_frames);

  tracker_params.max_trace_length = 10;
  nh.param("max_trace_length", tracker_params.max_trace_length, tracker_params.max_trace_length);

  tracker_ = std::unique_ptr<CTracker>(new CTracker(tracker_params));

  // setup dynamic reconfigure
  dynamic_recfg_ = new dynamic_reconfigure::Server<CostmapToDynamicObstaclesConfig>(nh);
  dynamic_reconfigure::Server<CostmapToDynamicObstaclesConfig>::CallbackType cb = boost::bind(&CostmapToDynamicObstacles::reconfigureCB, this, _1, _2);
  dynamic_recfg_->setCallback(cb);
}

void CostmapToDynamicObstacles::compute()
{
  if (costmap_mat_.empty())
    return;

  // Dynamic obstacles are separated from static obstacles
  int origin_x = round(costmap_->getOriginX() / costmap_->getResolution());
  int origin_y = round(costmap_->getOriginY() / costmap_->getResolution());
  // ROS_INFO("Origin x  [m]: %f    Origin_y  [m]: %f", costmap_->getOriginX(), costmap_->getOriginY());
  // ROS_INFO("Origin x [px]: %d \t Origin_y [px]: %d", originX, originY);

  bg_sub_->apply(costmap_mat_, fg_mask_, origin_x, origin_y);

  // if no foreground object is detected, no ObstacleMsgs need to be published
  if (fg_mask_.empty())
    return;

  cv::Mat bg_mat;
  if (publish_static_obstacles_)
  {
    // Get static obstacles
    bg_mat = costmap_mat_ - fg_mask_;
    // visualize("bg_mat", bg_mat);
  }


  // Centers and contours of Blobs are detected
  blob_det_->detect(fg_mask_, keypoints_);
  std::vector<std::vector<cv::Point>> contours = blob_det_->getContours();


  // Objects are assigned to objects from previous frame based on Hungarian Algorithm
  // Object velocities are estimated using a Kalman Filter
  std::vector<Point_t> detected_centers(keypoints_.size());
  for (int i = 0; i < keypoints_.size(); i++)
  {
    detected_centers.at(i).x = keypoints_.at(i).pt.x;
    detected_centers.at(i).y = keypoints_.at(i).pt.y;
    detected_centers.at(i).z = 0; // Currently unused!
  }

  tracker_->Update(detected_centers, contours);


  /*
  cv::Mat fg_mask_with_keypoints = cv::Mat::zeros(fg_mask.size(), CV_8UC3);
  cv::cvtColor(fg_mask, fg_mask_with_keypoints, cv::COLOR_GRAY2BGR);

  for (int i = 0; i < (int)tracker_->tracks.size(); ++i)
    cv::rectangle(fg_mask_with_keypoints, cv::boundingRect(tracker_->tracks[i]->getLastContour()),
                  cv::Scalar(0, 0, 255), 1);

  //visualize("fgMaskWithKeyPoints", fgMaskWithKeypoints);
  //cv::imwrite("/home/albers/Desktop/fgMask.png", fgMask);
  //cv::imwrite("/home/albers/Desktop/fgMaskWithKeyPoints.png", fgMaskWithKeypoints);
  */

  ObstacleArrayPtr obstacles(new ObstacleArrayMsg);
  // header.seq is automatically filled
  obstacles->header.stamp = ros::Time::now();
  obstacles->header.frame_id = "/map"; //Global frame /map

  // For all tracked objects
  for (unsigned int i = 0; i < (unsigned int)tracker_->tracks.size(); ++i)
  {
    geometry_msgs::Polygon polygon;

    // TODO directly create polygon inside getContour and avoid copy
    std::vector<Point_t> contour;
    getContour(i, contour); // this method also transforms map to world coordinates

    // convert contour to polygon
    for (const Point_t& pt : contour)
    {
      polygon.points.emplace_back();
      polygon.points.back().x = pt.x;
      polygon.points.back().y = pt.y;
      polygon.points.back().z = 0;
    }

    obstacles->obstacles.emplace_back();
    obstacles->obstacles.back().polygon = polygon;

    // Set obstacle ID
    obstacles->obstacles.back().id = tracker_->tracks.at(i)->track_id;

    // Set orientation
    geometry_msgs::QuaternionStamped orientation;

    Point_t vel = getEstimatedVelocityOfObject(i);
    double yaw = std::atan2(vel.y, vel.x);
    //ROS_INFO("yaw: %f", yaw);
    obstacles->obstacles.back().orientation = tf::createQuaternionMsgFromYaw(yaw);

    // Set velocity
    geometry_msgs::TwistWithCovariance velocities;
    //velocities.twist.linear.x = std::sqrt(vel.x*vel.x + vel.y*vel.y);
    //velocities.twist.linear.y = 0;
    velocities.twist.linear.x = vel.x;
    velocities.twist.linear.y = vel.y; // TODO(roesmann): don't we need to consider the transformation between opencv's and costmap's coordinate frames?
    velocities.twist.linear.z = 0;
    velocities.twist.angular.x = 0;
    velocities.twist.angular.y = 0;
    velocities.twist.angular.z = 0;

    // TODO: use correct covariance matrix
    velocities.covariance = {1, 0, 0, 0, 0, 0,
                             0, 1, 0, 0, 0, 0,
                             0, 0, 1, 0, 0, 0,
                             0, 0, 0, 1, 0, 0,
                             0, 0, 0, 0, 1, 0,
                             0, 0, 0, 0, 0, 1};

    obstacles->obstacles.back().velocities = velocities;
  }

  if (publish_static_obstacles_)
  {
    uchar* img_data = bg_mat.data;
    int width = bg_mat.cols;
    int height = bg_mat.rows;
    int stride = bg_mat.step;

    for(int i = 0; i < height; i++)
    {
      for(int j = 0; j < width; j++)
      {
          uchar value = img_data[i * stride + j];
          if (value > 0)
          {
            // obstacle found
            obstacles->obstacles.emplace_back();
            geometry_msgs::Point32 pt;
            pt.x = (double)j*costmap_->getResolution() + costmap_->getOriginX();
            pt.y = (double)i*costmap_->getResolution() + costmap_->getOriginY();
            obstacles->obstacles.back().polygon.points.push_back(pt);
            obstacles->obstacles.back().velocities.twist.linear.x = 0;
            obstacles->obstacles.back().velocities.twist.linear.y = 0;
            obstacles->obstacles.back().id = -1;
          }
      }
    }
  }

  updateObstacleContainer(obstacles);
}

void CostmapToDynamicObstacles::setCostmap2D(costmap_2d::Costmap2D* costmap)
{
  if (!costmap)
    return;

  costmap_ = costmap;

  updateCostmap2D();
}

void CostmapToDynamicObstacles::updateCostmap2D()
{
  if (!costmap_->getMutex())
  {
    ROS_ERROR("Cannot update costmap since the mutex pointer is null");
    return;
  }
  costmap_2d::Costmap2D::mutex_t::scoped_lock lock(*costmap_->getMutex());

  // Initialize costmapMat_ directly with the unsigned char array of costmap_
  //costmap_mat_ = cv::Mat(costmap_->getSizeInCellsX(), costmap_->getSizeInCellsY(), CV_8UC1,
  //                      costmap_->getCharMap()).clone(); // Deep copy: TODO
  costmap_mat_ = cv::Mat(costmap_->getSizeInCellsX(), costmap_->getSizeInCellsY(), CV_8UC1,
                        costmap_->getCharMap());
}

ObstacleArrayConstPtr CostmapToDynamicObstacles::getObstacles()
{
  boost::mutex::scoped_lock lock(mutex_);
  return obstacles_;
}

void CostmapToDynamicObstacles::updateObstacleContainer(ObstacleArrayPtr obstacles)
{
  boost::mutex::scoped_lock lock(mutex_);
  obstacles_ = obstacles;
}

Point_t CostmapToDynamicObstacles::getEstimatedVelocityOfObject(unsigned int idx)
{
  // vel [px/s] * costmapResolution [m/px] = vel [m/s]
  Point_t vel = tracker_->tracks.at(idx)->getEstimatedVelocity() * costmap_->getResolution() + ego_vel_;

  //ROS_INFO("vel x: %f, vel y: %f, vel z: %f", vel.x, vel.y, vel.z);
  // velocity in /map frame
  return vel;
}

void CostmapToDynamicObstacles::odomCallback(const nav_msgs::Odometry::ConstPtr& msg)
{
  ROS_INFO_ONCE("CostmapToDynamicObstacles: odom received.");

  tf::Quaternion pose;
  tf::quaternionMsgToTF(msg->pose.pose.orientation, pose);

  tf::Vector3 twistLinear;
  tf::vector3MsgToTF(msg->twist.twist.linear, twistLinear);

  // velocity of the robot in x, y and z coordinates
  tf::Vector3 vel = tf::quatRotate(pose, twistLinear);
  ego_vel_.x = vel.x();
  ego_vel_.y = vel.y();
  ego_vel_.z = vel.z();
}

void CostmapToDynamicObstacles::reconfigureCB(CostmapToDynamicObstaclesConfig& config, uint32_t level)
{
  publish_static_obstacles_ = config.publish_static_obstacles;

  // BackgroundSubtraction Parameters
  BackgroundSubtractor::Params bg_sub_params;
  bg_sub_params.alpha_slow = config.alpha_slow;
  bg_sub_params.alpha_fast = config.alpha_fast;
  bg_sub_params.beta = config.beta;
  bg_sub_params.min_sep_between_fast_and_slow_filter = config.min_sep_between_slow_and_fast_filter;
  bg_sub_params.min_occupancy_probability = config.min_occupancy_probability;
  bg_sub_params.max_occupancy_neighbors = config.max_occupancy_neighbors;
  bg_sub_params.morph_size = config.morph_size;
  bg_sub_->updateParameters(bg_sub_params);

  // BlobDetector Parameters
  BlobDetector::Params blob_det_params;
  // necessary, because blobDetParams are otherwise initialized with default values for dark blobs
  blob_det_params.filterByColor = true; // actually filterByIntensity, always true
  blob_det_params.blobColor = 255;      // Extract light blobs
  blob_det_params.thresholdStep = 256;  // Input for blobDetection is already a binary image
  blob_det_params.minThreshold = 127;
  blob_det_params.maxThreshold = 255;
  blob_det_params.minRepeatability = 1;
  blob_det_params.minDistBetweenBlobs = config.min_distance_between_blobs; // TODO: Currently not working as expected
  blob_det_params.filterByArea = config.filter_by_area;
  blob_det_params.minArea = config.min_area;
  blob_det_params.maxArea = config.max_area;
  blob_det_params.filterByCircularity = config.filter_by_circularity;
  blob_det_params.minCircularity = config.min_circularity;
  blob_det_params.maxCircularity = config.max_circularity;
  blob_det_params.filterByInertia = config.filter_by_inertia;
  blob_det_params.minInertiaRatio = config.min_inertia_ratio;
  blob_det_params.maxInertiaRatio = config.max_inertia_ratio;
  blob_det_params.filterByConvexity = config.filter_by_convexity;
  blob_det_params.minConvexity = config.min_convexity;
  blob_det_params.maxConvexity = config.max_convexity;
  blob_det_->updateParameters(blob_det_params);

  // Tracking Parameters
  CTracker::Params tracking_params;
  tracking_params.dt = config.dt;
  tracking_params.dist_thresh = config.dist_thresh;
  tracking_params.max_allowed_skipped_frames = config.max_allowed_skipped_frames;
  tracking_params.max_trace_length = config.max_trace_length;
  tracker_->updateParameters(tracking_params);
}

void CostmapToDynamicObstacles::getContour(unsigned int idx, std::vector<Point_t>& contour)
{
  assert(!tracker_->tracks.empty() && idx < tracker_->tracks.size());

  contour.clear();

  // contour [px] * costmapResolution [m/px] = contour [m]
  std::vector<cv::Point> contour2i = tracker_->tracks.at(idx)->getLastContour();

  contour.reserve(contour2i.size());

  Point_t costmap_origin(costmap_->getOriginX(), costmap_->getOriginY(), 0);

  for (std::size_t i = 0; i < contour2i.size(); ++i)
  {
    contour.push_back((Point_t(contour2i.at(i).x, contour2i.at(i).y, 0.0)*costmap_->getResolution())
                        + costmap_origin); // Shift to /map
  }

}

void CostmapToDynamicObstacles::visualize(const std::string& name, const cv::Mat& image)
{
  if (!image.empty())
  {
    cv::Mat im = image.clone();
    cv::flip(im, im, 0);
    cv::imshow(name, im);
    cv::waitKey(1);
  }
}
}