EGO-planner代码笔记1 grid_map.cpp源码解析

EGO-planner代码笔记1 grid_map.cpp源码解析

本帖持续更新，先把源码注释贴出来，在grid_map命名空间下的参数作用都能找得到

#include "plan_env/grid_map.h"

// #define current_img_ md_.depth_image_[image_cnt_ & 1]
// #define last_img_ md_.depth_image_[!(image_cnt_ & 1)]

void GridMap::initMap(ros::NodeHandle &nh)
{
  node_ = nh;

  /* get parameter */
  double x_size, y_size, z_size;
  //分辨率
  node_.param("grid_map/resolution", mp_.resolution_, -1.0);

  //地图大小(真实世界大小，不是栅格数量，单位m)
  node_.param("grid_map/map_size_x", x_size, -1.0);
  node_.param("grid_map/map_size_y", y_size, -1.0);
  node_.param("grid_map/map_size_z", z_size, -1.0);

  //局部更新范围
  node_.param("grid_map/local_update_range_x", mp_.local_update_range_(0), -1.0);
  node_.param("grid_map/local_update_range_y", mp_.local_update_range_(1), -1.0);
  node_.param("grid_map/local_update_range_z", mp_.local_update_range_(2), -1.0);

  //障碍物膨胀半径
  node_.param("grid_map/obstacles_inflation", mp_.obstacles_inflation_, -1.0);

  //相机内参
  node_.param("grid_map/fx", mp_.fx_, -1.0);
  node_.param("grid_map/fy", mp_.fy_, -1.0);
  node_.param("grid_map/cx", mp_.cx_, -1.0);
  node_.param("grid_map/cy", mp_.cy_, -1.0);

  //是否使用滤波器
  node_.param("grid_map/use_depth_filter", mp_.use_depth_filter_, true);
  //源码里注释掉了，没用
  node_.param("grid_map/depth_filter_tolerance", mp_.depth_filter_tolerance_, -1.0);

  //滤波器认为有效的最大距离和最小距离
  node_.param("grid_map/depth_filter_maxdist", mp_.depth_filter_maxdist_, -1.0);
  node_.param("grid_map/depth_filter_mindist", mp_.depth_filter_mindist_, -1.0);

  //抛弃深度相机的边缘
  node_.param("grid_map/depth_filter_margin", mp_.depth_filter_margin_, -1);

  //mm转化为m
  node_.param("grid_map/k_depth_scaling_factor", mp_.k_depth_scaling_factor_, -1.0);

  //降采样
  node_.param("grid_map/skip_pixel", mp_.skip_pixel_, -1);

  //概率栅格更新参数后续介绍
  node_.param("grid_map/p_hit", mp_.p_hit_, 0.70);
  node_.param("grid_map/p_miss", mp_.p_miss_, 0.35);
  node_.param("grid_map/p_min", mp_.p_min_, 0.12);
  node_.param("grid_map/p_max", mp_.p_max_, 0.97);
  node_.param("grid_map/p_occ", mp_.p_occ_, 0.80);

  //最长最短的深度相机光线长度
  node_.param("grid_map/min_ray_length", mp_.min_ray_length_, -0.1);
  node_.param("grid_map/max_ray_length", mp_.max_ray_length_, -0.1);

  //可视化截断高度，超过该高度的地图像素将不被显示
  node_.param("grid_map/visualization_truncate_height", mp_.visualization_truncate_height_, 999.0);

  //在一定高度上添加虚拟地面
  node_.param("grid_map/virtual_ceil_height", mp_.virtual_ceil_height_, -0.1);

  //源码里被注释了，没用
  node_.param("grid_map/show_occ_time", mp_.show_occ_time_, false);

  //接受的位姿1为geometry_msgs::Pose类型，2为nav_msgs::Odom类型
  node_.param("grid_map/pose_type", mp_.pose_type_, 1);

  //发布地图的frame_id
  node_.param("grid_map/frame_id", mp_.frame_id_, string("world"));

  //局部地图的边界，更新时以local_bound_min和local_bound_max向外扩展margin个栅格作为更新范围
  node_.param("grid_map/local_map_margin", mp_.local_map_margin_, 1);

  //地图坐标系下的地面高度
  node_.param("grid_map/ground_height", mp_.ground_height_, 1.0);

  //地图原点和地图大小(m)
  mp_.resolution_inv_ = 1 / mp_.resolution_;
  mp_.map_origin_ = Eigen::Vector3d(-x_size / 2.0, -y_size / 2.0, mp_.ground_height_);
  mp_.map_size_ = Eigen::Vector3d(x_size, y_size, z_size);

  //#define logit(x) (log((x) / (1 - (x)))),贝叶斯滤波相关，后续介绍
  mp_.prob_hit_log_ = logit(mp_.p_hit_);
  mp_.prob_miss_log_ = logit(mp_.p_miss_);
  mp_.clamp_min_log_ = logit(mp_.p_min_);
  mp_.clamp_max_log_ = logit(mp_.p_max_);
  mp_.min_occupancy_log_ = logit(mp_.p_occ_);
  mp_.unknown_flag_ = 0.01;

  cout << "hit: " << mp_.prob_hit_log_ << endl;
  cout << "miss: " << mp_.prob_miss_log_ << endl;
  cout << "min log: " << mp_.clamp_min_log_ << endl;
  cout << "max: " << mp_.clamp_max_log_ << endl;
  cout << "thresh log: " << mp_.min_occupancy_log_ << endl;

  for (int i = 0; i < 3; ++i)
    mp_.map_voxel_num_(i) = ceil(mp_.map_size_(i) / mp_.resolution_);
  //地图的最大和最小边界(m)
  mp_.map_min_boundary_ = mp_.map_origin_;
  mp_.map_max_boundary_ = mp_.map_origin_ + mp_.map_size_;

  // initialize data buffers

  int buffer_size = mp_.map_voxel_num_(0) * mp_.map_voxel_num_(1) * mp_.map_voxel_num_(2);
  //概率占据栅格地图，以及膨胀后的占据栅格地图
  md_.occupancy_buffer_ = vector<double>(buffer_size, mp_.clamp_min_log_ - mp_.unknown_flag_);
  md_.occupancy_buffer_inflate_ = vector<char>(buffer_size, 0);

  //命中和通过某栅格的计数器count_hit_and_miss_ - count_hit_ 即为miss数
  md_.count_hit_and_miss_ = vector<short>(buffer_size, 0);
  md_.count_hit_ = vector<short>(buffer_size, 0);
  
  //标记某栅格是第几帧的终点，如果是同一帧，raycast的效果是一样的，不会进行同样操作
  md_.flag_rayend_ = vector<char>(buffer_size, -1);

  //标记某栅格是第几帧的透过点
  md_.flag_traverse_ = vector<char>(buffer_size, -1);

  // raycast轮次和上面flag_rayend_一起使用
  md_.raycast_num_ = 0;

  //深度图转点云的存储位置
  md_.proj_points_.resize(640 * 480 / mp_.skip_pixel_ / mp_.skip_pixel_);
  md_.proj_points_cnt = 0;

  //相机和机体的坐标转换
  md_.cam2body_ << 0.0, 0.0, 1.0, 0.0,
      -1.0, 0.0, 0.0, 0.0,
      0.0, -1.0, 0.0, -0.02,
      0.0, 0.0, 0.0, 1.0;

  /* init callback */
  //深度图像和位姿时间戳对齐
  depth_sub_.reset(new message_filters::Subscriber<sensor_msgs::Image>(node_, "/grid_map/depth", 50));

  if (mp_.pose_type_ == POSE_STAMPED)
  {
    pose_sub_.reset(
        new message_filters::Subscriber<geometry_msgs::PoseStamped>(node_, "/grid_map/pose", 25));

    sync_image_pose_.reset(new message_filters::Synchronizer<SyncPolicyImagePose>(
        SyncPolicyImagePose(100), *depth_sub_, *pose_sub_));
    sync_image_pose_->registerCallback(boost::bind(&GridMap::depthPoseCallback, this, _1, _2));
  }
  else if (mp_.pose_type_ == ODOMETRY)
  {
    odom_sub_.reset(new message_filters::Subscriber<nav_msgs::Odometry>(node_, "/grid_map/odom", 100));

    sync_image_odom_.reset(new message_filters::Synchronizer<SyncPolicyImageOdom>(
        SyncPolicyImageOdom(100), *depth_sub_, *odom_sub_));
    sync_image_odom_->registerCallback(boost::bind(&GridMap::depthOdomCallback, this, _1, _2));
  }
  //使用深度图的话会自动禁止下面的
  // use odometry and point cloud
  indep_cloud_sub_ =
      node_.subscribe<sensor_msgs::PointCloud2>("/grid_map/cloud", 10, &GridMap::cloudCallback, this);
  indep_odom_sub_ =
      node_.subscribe<nav_msgs::Odometry>("/grid_map/odom", 10, &GridMap::odomCallback, this);

  //地图更新和可视化都是通过定时器回调实现
  occ_timer_ = node_.createTimer(ros::Duration(0.05), &GridMap::updateOccupancyCallback, this);
  vis_timer_ = node_.createTimer(ros::Duration(0.05), &GridMap::visCallback, this);

  //概率地图和膨胀地图
  map_pub_ = node_.advertise<sensor_msgs::PointCloud2>("/grid_map/occupancy", 10);
  map_inf_pub_ = node_.advertise<sensor_msgs::PointCloud2>("/grid_map/occupancy_inflate", 10);

  //发布未知点云，源码中实际并未使用
  unknown_pub_ = node_.advertise<sensor_msgs::PointCloud2>("/grid_map/unknown", 10);

  //有新数据，举要更新
  md_.occ_need_update_ = false;
  //raycast后需要重新计算概率
  md_.local_updated_ = false;

  //是否有第一帧深度图，如果有则odomCallback不生效
  md_.has_first_depth_ = false;

  //是否有odom，没有会把点云回调直接return
  md_.has_odom_ = false;
  //是否有点云，目前没用
  md_.has_cloud_ = false;
  //注释了，没用
  md_.image_cnt_ = 0;
  //分析时间的没用
  md_.fuse_time_ = 0.0;

  //这个参数没用到
  md_.update_num_ = 0;
  md_.max_fuse_time_ = 0.0;

  // rand_noise_ = uniform_real_distribution<double>(-0.2, 0.2);
  // rand_noise2_ = normal_distribution<double>(0, 0.2);
  // random_device rd;
  // eng_ = default_random_engine(rd());
}

//重置地图occupancy_buffer_inflate_
void GridMap::resetBuffer()
{
  Eigen::Vector3d min_pos = mp_.map_min_boundary_;
  Eigen::Vector3d max_pos = mp_.map_max_boundary_;

  resetBuffer(min_pos, max_pos);

  md_.local_bound_min_ = Eigen::Vector3i::Zero();
  md_.local_bound_max_ = mp_.map_voxel_num_ - Eigen::Vector3i::Ones();
}

void GridMap::resetBuffer(Eigen::Vector3d min_pos, Eigen::Vector3d max_pos)
{

  Eigen::Vector3i min_id, max_id;
  posToIndex(min_pos, min_id);
  posToIndex(max_pos, max_id);

  boundIndex(min_id);
  boundIndex(max_id);

  /* reset occ and dist buffer */
  for (int x = min_id(0); x <= max_id(0); ++x)
    for (int y = min_id(1); y <= max_id(1); ++y)
      for (int z = min_id(2); z <= max_id(2); ++z)
      {
        md_.occupancy_buffer_inflate_[toAddress(x, y, z)] = 0;
      }
}

//记录某个位置的击中和通过，并缓存，返回该位置的储存地址
int GridMap::setCacheOccupancy(Eigen::Vector3d pos, int occ)
{
  if (occ != 1 && occ != 0)
    return INVALID_IDX;

  Eigen::Vector3i id;
  posToIndex(pos, id);
  int idx_ctns = toAddress(id);

  md_.count_hit_and_miss_[idx_ctns] += 1;

  //初次被通过或被击中时加入缓存队列
  if (md_.count_hit_and_miss_[idx_ctns] == 1)
  {
    md_.cache_voxel_.push(id);
  }

  //击中次数+1
  if (occ == 1)
    md_.count_hit_[idx_ctns] += 1;

  return idx_ctns;
}

//处理深度图
void GridMap::projectDepthImage()
{
  // md_.proj_points_.clear();
  md_.proj_points_cnt = 0;

  uint16_t *row_ptr;
  // int cols = current_img_.cols, rows = current_img_.rows;
  int cols = md_.depth_image_.cols;
  int rows = md_.depth_image_.rows;

  double depth;

  //相机位姿的四元数转旋转矩阵
  Eigen::Matrix3d camera_r = md_.camera_q_.toRotationMatrix();

  // cout << "rotate: " << md_.camera_q_.toRotationMatrix() << endl;
  // std::cout << "pos in proj: " << md_.camera_pos_ << std::endl;

  //是否使用深度滤波
  if (!mp_.use_depth_filter_)
  {
    for (int v = 0; v < rows; v++)
    {
      row_ptr = md_.depth_image_.ptr<uint16_t>(v);

      for (int u = 0; u < cols; u++)
      {

        Eigen::Vector3d proj_pt;
        depth = (*row_ptr++) / mp_.k_depth_scaling_factor_;
        proj_pt(0) = (u - mp_.cx_) * depth / mp_.fx_;
        proj_pt(1) = (v - mp_.cy_) * depth / mp_.fy_;
        proj_pt(2) = depth;

        proj_pt = camera_r * proj_pt + md_.camera_pos_;

        if (u == 320 && v == 240)
          std::cout << "depth: " << depth << std::endl;
        md_.proj_points_[md_.proj_points_cnt++] = proj_pt;
      }
    }
  }
  /* use depth filter */
  else
  {
    //是否有第一帧深度图
    if (!md_.has_first_depth_)
      md_.has_first_depth_ = true;
    else
    {
      Eigen::Vector3d pt_cur, pt_world, pt_reproj;

      //上一帧旋转的逆矩阵
      Eigen::Matrix3d last_camera_r_inv;
      last_camera_r_inv = md_.last_camera_q_.inverse();
      const double inv_factor = 1.0 / mp_.k_depth_scaling_factor_;

      //去掉边缘，以skip_pixel_降采样
      for (int v = mp_.depth_filter_margin_; v < rows - mp_.depth_filter_margin_; v += mp_.skip_pixel_)
      {
        row_ptr = md_.depth_image_.ptr<uint16_t>(v) + mp_.depth_filter_margin_;

        for (int u = mp_.depth_filter_margin_; u < cols - mp_.depth_filter_margin_;
             u += mp_.skip_pixel_)
        {

          depth = (*row_ptr) * inv_factor;
          row_ptr = row_ptr + mp_.skip_pixel_;

          //添加噪音
          // filter depth
          // depth += rand_noise_(eng_);
          // if (depth > 0.01) depth += rand_noise2_(eng_);

          //无效点赋最大值
          if (*row_ptr == 0)
          {
            depth = mp_.max_ray_length_ + 0.1;
          }
          else if (depth < mp_.depth_filter_mindist_)
          {
            //低于滤波器最小值跳过，不会被标记
            continue;
          }
          else if (depth > mp_.depth_filter_maxdist_)
          {
            //大于滤波器最大值直接赋最大值
            depth = mp_.max_ray_length_ + 0.1;
          }
          //相机坐标系下的点
          // project to world frame
          pt_cur(0) = (u - mp_.cx_) * depth / mp_.fx_;
          pt_cur(1) = (v - mp_.cy_) * depth / mp_.fy_;
          pt_cur(2) = depth;
          //世界坐标下的点
          pt_world = camera_r * pt_cur + md_.camera_pos_;
          // if (!isInMap(pt_world)) {
          //   pt_world = closetPointInMap(pt_world, md_.camera_pos_);
          // }
          //存起来后续处理
          md_.proj_points_[md_.proj_points_cnt++] = pt_world;

          //检查两帧之间的连续性，没有使用
          // check consistency with last image, disabled...
          if (false)
          {
            pt_reproj = last_camera_r_inv * (pt_world - md_.last_camera_pos_);
            double uu = pt_reproj.x() * mp_.fx_ / pt_reproj.z() + mp_.cx_;
            double vv = pt_reproj.y() * mp_.fy_ / pt_reproj.z() + mp_.cy_;

            if (uu >= 0 && uu < cols && vv >= 0 && vv < rows)
            {
              if (fabs(md_.last_depth_image_.at<uint16_t>((int)vv, (int)uu) * inv_factor -
                       pt_reproj.z()) < mp_.depth_filter_tolerance_)
              {
                md_.proj_points_[md_.proj_points_cnt++] = pt_world;
              }
            }
            else
            {
              md_.proj_points_[md_.proj_points_cnt++] = pt_world;
            }
          }
        }
      }
    }
  }

  /* maintain camera pose for consistency check */

  md_.last_camera_pos_ = md_.camera_pos_;
  md_.last_camera_q_ = md_.camera_q_;
  md_.last_depth_image_ = md_.depth_image_;
}

//光线投射，沿着光路更新路径上的所有栅格概率
void GridMap::raycastProcess()
{
  // if (md_.proj_points_.size() == 0)
  if (md_.proj_points_cnt == 0)
    return;

  ros::Time t1, t2;

  md_.raycast_num_ += 1;

  int vox_idx;
  double length;

  //地图更新的边界，先把上界设置为最小值，下界设置为最大值
  // bounding box of updated region
  double min_x = mp_.map_max_boundary_(0);
  double min_y = mp_.map_max_boundary_(1);
  double min_z = mp_.map_max_boundary_(2);

  double max_x = mp_.map_min_boundary_(0);
  double max_y = mp_.map_min_boundary_(1);
  double max_z = mp_.map_min_boundary_(2);

  //光线投射器
  RayCaster raycaster;
  Eigen::Vector3d half = Eigen::Vector3d(0.5, 0.5, 0.5);
  Eigen::Vector3d ray_pt, pt_w;

  for (int i = 0; i < md_.proj_points_cnt; ++i)
  {
    pt_w = md_.proj_points_[i];

    // set flag for projected point

    if (!isInMap(pt_w))
    {
      //不在地图里则找到地图边界点
      pt_w = closetPointInMap(pt_w, md_.camera_pos_);
      //长度超过最大光线追踪长度则取最大追踪长度
      length = (pt_w - md_.camera_pos_).norm();
      if (length > mp_.max_ray_length_)
      {
        pt_w = (pt_w - md_.camera_pos_) / length * mp_.max_ray_length_ + md_.camera_pos_;
      }
      //末端标记为miss
      vox_idx = setCacheOccupancy(pt_w, 0);
    }
    else
    {
      //超过最长光线追踪长度会在末端标记为miss，没超过会在末端标记为hit
      length = (pt_w - md_.camera_pos_).norm();

      if (length > mp_.max_ray_length_)
      {
        pt_w = (pt_w - md_.camera_pos_) / length * mp_.max_ray_length_ + md_.camera_pos_;
        vox_idx = setCacheOccupancy(pt_w, 0);
      }
      else
      {
        vox_idx = setCacheOccupancy(pt_w, 1);
      }
    }

    max_x = max(max_x, pt_w(0));
    max_y = max(max_y, pt_w(1));
    max_z = max(max_z, pt_w(2));

    min_x = min(min_x, pt_w(0));
    min_y = min(min_y, pt_w(1));
    min_z = min(min_z, pt_w(2));

    // raycasting between camera center and point
    //标记某栅格是第几帧的终点，如果是同一帧，raycast的效果是一样的，不会进行同样操作
    if (vox_idx != INVALID_IDX)
    {
      if (md_.flag_rayend_[vox_idx] == md_.raycast_num_)
      {
        continue;
      }
      else
      {
        md_.flag_rayend_[vox_idx] = md_.raycast_num_;
      }
    }

    raycaster.setInput(pt_w / mp_.resolution_, md_.camera_pos_ / mp_.resolution_);

    while (raycaster.step(ray_pt))
    {
      Eigen::Vector3d tmp = (ray_pt + half) * mp_.resolution_;
      length = (tmp - md_.camera_pos_).norm();

      // if (length < mp_.min_ray_length_) break;

      vox_idx = setCacheOccupancy(tmp, 0);

      //这里break很巧妙，第一次看一以为应该是continue，但是我们注意到，racast的操作是从点云向相机反向进行的，
      //并且相机的所有raycast都会汇聚到同一个点，当raycast进行到某处发现该帧已经被其他点raycast过了，
      //则可以保证后续所有的点都是raycast过的
      if (vox_idx != INVALID_IDX)
      {
        if (md_.flag_traverse_[vox_idx] == md_.raycast_num_)
        {
          break;
        }
        else
        {
          md_.flag_traverse_[vox_idx] = md_.raycast_num_;
        }
      }
    }
  }

  min_x = min(min_x, md_.camera_pos_(0));
  min_y = min(min_y, md_.camera_pos_(1));
  min_z = min(min_z, md_.camera_pos_(2));

  max_x = max(max_x, md_.camera_pos_(0));
  max_y = max(max_y, md_.camera_pos_(1));
  max_z = max(max_z, md_.camera_pos_(2));
  max_z = max(max_z, mp_.ground_height_);

  //所有raycast涉及到的边界
  posToIndex(Eigen::Vector3d(max_x, max_y, max_z), md_.local_bound_max_);
  posToIndex(Eigen::Vector3d(min_x, min_y, min_z), md_.local_bound_min_);
  boundIndex(md_.local_bound_min_);
  boundIndex(md_.local_bound_max_);

  md_.local_updated_ = true;

  // update occupancy cached in queue
  //局部地图更新范围限制在全局地图内
  Eigen::Vector3d local_range_min = md_.camera_pos_ - mp_.local_update_range_;
  Eigen::Vector3d local_range_max = md_.camera_pos_ + mp_.local_update_range_;

  Eigen::Vector3i min_id, max_id;
  posToIndex(local_range_min, min_id);
  posToIndex(local_range_max, max_id);
  boundIndex(min_id);
  boundIndex(max_id);

  // std::cout << "cache all: " << md_.cache_voxel_.size() << std::endl;

  while (!md_.cache_voxel_.empty())
  {

    Eigen::Vector3i idx = md_.cache_voxel_.front();
    int idx_ctns = toAddress(idx);
    md_.cache_voxel_.pop();

    //命中的多就按命中更新，通过的多就按通过更新
    double log_odds_update =
        md_.count_hit_[idx_ctns] >= md_.count_hit_and_miss_[idx_ctns] - md_.count_hit_[idx_ctns] ? mp_.prob_hit_log_ : mp_.prob_miss_log_;

    md_.count_hit_[idx_ctns] = md_.count_hit_and_miss_[idx_ctns] = 0;

    if (log_odds_update >= 0 && md_.occupancy_buffer_[idx_ctns] >= mp_.clamp_max_log_)
    {
    //命中大于log概率clamp_max_log_的不更新
      continue;
    }
    else if (log_odds_update <= 0 && md_.occupancy_buffer_[idx_ctns] <= mp_.clamp_min_log_)
    {
    //命中小于log概率clamp_min_log_的更新为最小值
      md_.occupancy_buffer_[idx_ctns] = mp_.clamp_min_log_;
      continue;
    }

    //判断是否在更新范围，不在的先改为最小值
    bool in_local = idx(0) >= min_id(0) && idx(0) <= max_id(0) && idx(1) >= min_id(1) &&
                    idx(1) <= max_id(1) && idx(2) >= min_id(2) && idx(2) <= max_id(2);
    if (!in_local)
    {
      md_.occupancy_buffer_[idx_ctns] = mp_.clamp_min_log_;
    }

    md_.occupancy_buffer_[idx_ctns] =
        std::min(std::max(md_.occupancy_buffer_[idx_ctns] + log_odds_update, mp_.clamp_min_log_),
                 mp_.clamp_max_log_);
  }
}

//找到一个地图外的点和相机连线和地图边界的交点
Eigen::Vector3d GridMap::closetPointInMap(const Eigen::Vector3d &pt, const Eigen::Vector3d &camera_pt)
{
  Eigen::Vector3d diff = pt - camera_pt;
  Eigen::Vector3d max_tc = mp_.map_max_boundary_ - camera_pt;
  Eigen::Vector3d min_tc = mp_.map_min_boundary_ - camera_pt;

  double min_t = 1000000;

  for (int i = 0; i < 3; ++i)
  {
    if (fabs(diff[i]) > 0)
    {

      double t1 = max_tc[i] / diff[i];
      if (t1 > 0 && t1 < min_t)
        min_t = t1;

      double t2 = min_tc[i] / diff[i];
      if (t2 > 0 && t2 < min_t)
        min_t = t2;
    }
  }

  return camera_pt + (min_t - 1e-3) * diff;
}

void GridMap::clearAndInflateLocalMap()
{
  /*clear outside local*/
  const int vec_margin = 5;
  // Eigen::Vector3i min_vec_margin = min_vec - Eigen::Vector3i(vec_margin,
  // vec_margin, vec_margin); Eigen::Vector3i max_vec_margin = max_vec +
  // Eigen::Vector3i(vec_margin, vec_margin, vec_margin);

  Eigen::Vector3i min_cut = md_.local_bound_min_ -
                            Eigen::Vector3i(mp_.local_map_margin_, mp_.local_map_margin_, mp_.local_map_margin_);
  Eigen::Vector3i max_cut = md_.local_bound_max_ +
                            Eigen::Vector3i(mp_.local_map_margin_, mp_.local_map_margin_, mp_.local_map_margin_);
  boundIndex(min_cut);
  boundIndex(max_cut);

  Eigen::Vector3i min_cut_m = min_cut - Eigen::Vector3i(vec_margin, vec_margin, vec_margin);
  Eigen::Vector3i max_cut_m = max_cut + Eigen::Vector3i(vec_margin, vec_margin, vec_margin);
  boundIndex(min_cut_m);
  boundIndex(max_cut_m);

  // clear data outside the local range
  //清除局部更新范围外的障碍，
  for (int x = min_cut_m(0); x <= max_cut_m(0); ++x)
    for (int y = min_cut_m(1); y <= max_cut_m(1); ++y)
    {

      for (int z = min_cut_m(2); z < min_cut(2); ++z)
      {
        int idx = toAddress(x, y, z);
        md_.occupancy_buffer_[idx] = mp_.clamp_min_log_ - mp_.unknown_flag_;
      }

      for (int z = max_cut(2) + 1; z <= max_cut_m(2); ++z)
      {
        int idx = toAddress(x, y, z);
        md_.occupancy_buffer_[idx] = mp_.clamp_min_log_ - mp_.unknown_flag_;
      }
    }

  for (int z = min_cut_m(2); z <= max_cut_m(2); ++z)
    for (int x = min_cut_m(0); x <= max_cut_m(0); ++x)
    {

      for (int y = min_cut_m(1); y < min_cut(1); ++y)
      {
        int idx = toAddress(x, y, z);
        md_.occupancy_buffer_[idx] = mp_.clamp_min_log_ - mp_.unknown_flag_;
      }

      for (int y = max_cut(1) + 1; y <= max_cut_m(1); ++y)
      {
        int idx = toAddress(x, y, z);
        md_.occupancy_buffer_[idx] = mp_.clamp_min_log_ - mp_.unknown_flag_;
      }
    }

  for (int y = min_cut_m(1); y <= max_cut_m(1); ++y)
    for (int z = min_cut_m(2); z <= max_cut_m(2); ++z)
    {

      for (int x = min_cut_m(0); x < min_cut(0); ++x)
      {
        int idx = toAddress(x, y, z);
        md_.occupancy_buffer_[idx] = mp_.clamp_min_log_ - mp_.unknown_flag_;
      }

      for (int x = max_cut(0) + 1; x <= max_cut_m(0); ++x)
      {
        int idx = toAddress(x, y, z);
        md_.occupancy_buffer_[idx] = mp_.clamp_min_log_ - mp_.unknown_flag_;
      }
    }

  // inflate occupied voxels to compensate robot size

  int inf_step = ceil(mp_.obstacles_inflation_ / mp_.resolution_);
  // int inf_step_z = 1;
  vector<Eigen::Vector3i> inf_pts(pow(2 * inf_step + 1, 3));
  // inf_pts.resize(4 * inf_step + 3);
  Eigen::Vector3i inf_pt;

  //清除局部更新范围内的障碍
  // clear outdated data
  for (int x = md_.local_bound_min_(0); x <= md_.local_bound_max_(0); ++x)
    for (int y = md_.local_bound_min_(1); y <= md_.local_bound_max_(1); ++y)
      for (int z = md_.local_bound_min_(2); z <= md_.local_bound_max_(2); ++z)
      {
        md_.occupancy_buffer_inflate_[toAddress(x, y, z)] = 0;
      }

  //膨胀局部更新范围内的障碍
  // inflate obstacles
  for (int x = md_.local_bound_min_(0); x <= md_.local_bound_max_(0); ++x)
    for (int y = md_.local_bound_min_(1); y <= md_.local_bound_max_(1); ++y)
      for (int z = md_.local_bound_min_(2); z <= md_.local_bound_max_(2); ++z)
      {
        //如果被占据
        if (md_.occupancy_buffer_[toAddress(x, y, z)] > mp_.min_occupancy_log_)
        {
          //先获取膨胀栅格，再遍历每个栅格赋值
          inflatePoint(Eigen::Vector3i(x, y, z), inf_step, inf_pts);

          for (int k = 0; k < (int)inf_pts.size(); ++k)
          {
            inf_pt = inf_pts[k];
            int idx_inf = toAddress(inf_pt);
            if (idx_inf < 0 ||
                idx_inf >= mp_.map_voxel_num_(0) * mp_.map_voxel_num_(1) * mp_.map_voxel_num_(2))
            {
              continue;
            }
            md_.occupancy_buffer_inflate_[idx_inf] = 1;
          }
        }
      }

  //添加一个虚拟地面
  // add virtual ceiling to limit flight height
  if (mp_.virtual_ceil_height_ > -0.5)
  {
    int ceil_id = floor((mp_.virtual_ceil_height_ - mp_.map_origin_(2)) * mp_.resolution_inv_);
    for (int x = md_.local_bound_min_(0); x <= md_.local_bound_max_(0); ++x)
      for (int y = md_.local_bound_min_(1); y <= md_.local_bound_max_(1); ++y)
      {
        md_.occupancy_buffer_inflate_[toAddress(x, y, ceil_id)] = 1;
      }
  }
}

//定期发布可视化的地图
void GridMap::visCallback(const ros::TimerEvent & /*event*/)
{

  publishMap();
  publishMapInflate(true);
}

//定期更新地图
void GridMap::updateOccupancyCallback(const ros::TimerEvent & /*event*/)
{
  if (!md_.occ_need_update_)
    return;

  /* update occupancy */
  // ros::Time t1, t2, t3, t4;
  // t1 = ros::Time::now();

  projectDepthImage();
  // t2 = ros::Time::now();
  raycastProcess();
  // t3 = ros::Time::now();

  if (md_.local_updated_)
    clearAndInflateLocalMap();

  // t4 = ros::Time::now();

  // cout << setprecision(7);
  // cout << "t2=" << (t2-t1).toSec() << " t3=" << (t3-t2).toSec() << " t4=" << (t4-t3).toSec() << endl;;

  // md_.fuse_time_ += (t2 - t1).toSec();
  // md_.max_fuse_time_ = max(md_.max_fuse_time_, (t2 - t1).toSec());

  // if (mp_.show_occ_time_)
  //   ROS_WARN("Fusion: cur t = %lf, avg t = %lf, max t = %lf", (t2 - t1).toSec(),
  //            md_.fuse_time_ / md_.update_num_, md_.max_fuse_time_);

  md_.occ_need_update_ = false;
  md_.local_updated_ = false;
}

//深度图和位姿的联合回调函数
void GridMap::depthPoseCallback(const sensor_msgs::ImageConstPtr &img,
                                const geometry_msgs::PoseStampedConstPtr &pose)
{
  /* get depth image */
  cv_bridge::CvImagePtr cv_ptr;
  cv_ptr = cv_bridge::toCvCopy(img, img->encoding);

  if (img->encoding == sensor_msgs::image_encodings::TYPE_32FC1)
  {
    (cv_ptr->image).convertTo(cv_ptr->image, CV_16UC1, mp_.k_depth_scaling_factor_);
  }
  cv_ptr->image.copyTo(md_.depth_image_);

  // std::cout << "depth: " << md_.depth_image_.cols << ", " << md_.depth_image_.rows << std::endl;

  /* get pose */
  md_.camera_pos_(0) = pose->pose.position.x;
  md_.camera_pos_(1) = pose->pose.position.y;
  md_.camera_pos_(2) = pose->pose.position.z;
  md_.camera_q_ = Eigen::Quaterniond(pose->pose.orientation.w, pose->pose.orientation.x,
                                     pose->pose.orientation.y, pose->pose.orientation.z);

  //相机不在地图里了没法更新
  if (isInMap(md_.camera_pos_))
  {
    md_.has_odom_ = true;
    md_.update_num_ += 1;
    md_.occ_need_update_ = true;
  }
  else
  {
    md_.occ_need_update_ = false;
  }
}

//odom回调，使用深度图则直接返回
void GridMap::odomCallback(const nav_msgs::OdometryConstPtr &odom)
{
  if (md_.has_first_depth_)
    return;

  md_.camera_pos_(0) = odom->pose.pose.position.x;
  md_.camera_pos_(1) = odom->pose.pose.position.y;
  md_.camera_pos_(2) = odom->pose.pose.position.z;

  md_.has_odom_ = true;
}

//点云回调和odom一起使用
void GridMap::cloudCallback(const sensor_msgs::PointCloud2ConstPtr &img)
{

  pcl::PointCloud<pcl::PointXYZ> latest_cloud;
  pcl::fromROSMsg(*img, latest_cloud);

  md_.has_cloud_ = true;

  if (!md_.has_odom_)
  {
    std::cout << "no odom!" << std::endl;
    return;
  }

  if (latest_cloud.points.size() == 0)
    return;

  if (isnan(md_.camera_pos_(0)) || isnan(md_.camera_pos_(1)) || isnan(md_.camera_pos_(2)))
    return;

  //清除更新范围内的点云
  this->resetBuffer(md_.camera_pos_ - mp_.local_update_range_,
                    md_.camera_pos_ + mp_.local_update_range_);

  pcl::PointXYZ pt;
  Eigen::Vector3d p3d, p3d_inf;

  int inf_step = ceil(mp_.obstacles_inflation_ / mp_.resolution_);
  int inf_step_z = 1;

  double max_x, max_y, max_z, min_x, min_y, min_z;

  min_x = mp_.map_max_boundary_(0);
  min_y = mp_.map_max_boundary_(1);
  min_z = mp_.map_max_boundary_(2);

  max_x = mp_.map_min_boundary_(0);
  max_y = mp_.map_min_boundary_(1);
  max_z = mp_.map_min_boundary_(2);

  for (size_t i = 0; i < latest_cloud.points.size(); ++i)
  {
    pt = latest_cloud.points[i];
    p3d(0) = pt.x, p3d(1) = pt.y, p3d(2) = pt.z;

    /* point inside update range */
    Eigen::Vector3d devi = p3d - md_.camera_pos_;
    Eigen::Vector3i inf_pt;
    //在局部更新范围内的点就直接膨胀然后更新上去，不算概率了
    if (fabs(devi(0)) < mp_.local_update_range_(0) && fabs(devi(1)) < mp_.local_update_range_(1) &&
        fabs(devi(2)) < mp_.local_update_range_(2))
    {

      /* inflate the point */
      for (int x = -inf_step; x <= inf_step; ++x)
        for (int y = -inf_step; y <= inf_step; ++y)
          for (int z = -inf_step_z; z <= inf_step_z; ++z)
          {

            p3d_inf(0) = pt.x + x * mp_.resolution_;
            p3d_inf(1) = pt.y + y * mp_.resolution_;
            p3d_inf(2) = pt.z + z * mp_.resolution_;

            max_x = max(max_x, p3d_inf(0));
            max_y = max(max_y, p3d_inf(1));
            max_z = max(max_z, p3d_inf(2));

            min_x = min(min_x, p3d_inf(0));
            min_y = min(min_y, p3d_inf(1));
            min_z = min(min_z, p3d_inf(2));

            posToIndex(p3d_inf, inf_pt);

            if (!isInMap(inf_pt))
              continue;

            int idx_inf = toAddress(inf_pt);

            md_.occupancy_buffer_inflate_[idx_inf] = 1;
          }
    }
  }

  min_x = min(min_x, md_.camera_pos_(0));
  min_y = min(min_y, md_.camera_pos_(1));
  min_z = min(min_z, md_.camera_pos_(2));

  max_x = max(max_x, md_.camera_pos_(0));
  max_y = max(max_y, md_.camera_pos_(1));
  max_z = max(max_z, md_.camera_pos_(2));

  max_z = max(max_z, mp_.ground_height_);

  posToIndex(Eigen::Vector3d(max_x, max_y, max_z), md_.local_bound_max_);
  posToIndex(Eigen::Vector3d(min_x, min_y, min_z), md_.local_bound_min_);

  boundIndex(md_.local_bound_min_);
  boundIndex(md_.local_bound_max_);
}

//发布地图点云
void GridMap::publishMap()
{

  if (map_pub_.getNumSubscribers() <= 0)
    return;

  pcl::PointXYZ pt;
  pcl::PointCloud<pcl::PointXYZ> cloud;
  //局部更新范围内的点云发布
  Eigen::Vector3i min_cut = md_.local_bound_min_;
  Eigen::Vector3i max_cut = md_.local_bound_max_;

  int lmm = mp_.local_map_margin_ / 2;
  min_cut -= Eigen::Vector3i(lmm, lmm, lmm);
  max_cut += Eigen::Vector3i(lmm, lmm, lmm);

  boundIndex(min_cut);
  boundIndex(max_cut);

  for (int x = min_cut(0); x <= max_cut(0); ++x)
    for (int y = min_cut(1); y <= max_cut(1); ++y)
      for (int z = min_cut(2); z <= max_cut(2); ++z)
      {
        if (md_.occupancy_buffer_[toAddress(x, y, z)] < mp_.min_occupancy_log_)
          continue;

        Eigen::Vector3d pos;
        indexToPos(Eigen::Vector3i(x, y, z), pos);
        if (pos(2) > mp_.visualization_truncate_height_)
          continue;
        pt.x = pos(0);
        pt.y = pos(1);
        pt.z = pos(2);
        cloud.push_back(pt);
      }

  cloud.width = cloud.points.size();
  cloud.height = 1;
  cloud.is_dense = true;
  cloud.header.frame_id = mp_.frame_id_;
  sensor_msgs::PointCloud2 cloud_msg;

  pcl::toROSMsg(cloud, cloud_msg);
  map_pub_.publish(cloud_msg);
}

void GridMap::publishMapInflate(bool all_info)
{

  if (map_inf_pub_.getNumSubscribers() <= 0)
    return;

  pcl::PointXYZ pt;
  pcl::PointCloud<pcl::PointXYZ> cloud;

  Eigen::Vector3i min_cut = md_.local_bound_min_;
  Eigen::Vector3i max_cut = md_.local_bound_max_;

  if (all_info)
  {
    int lmm = mp_.local_map_margin_;
    min_cut -= Eigen::Vector3i(lmm, lmm, lmm);
    max_cut += Eigen::Vector3i(lmm, lmm, lmm);
  }

  boundIndex(min_cut);
  boundIndex(max_cut);

  for (int x = min_cut(0); x <= max_cut(0); ++x)
    for (int y = min_cut(1); y <= max_cut(1); ++y)
      for (int z = min_cut(2); z <= max_cut(2); ++z)
      {
        if (md_.occupancy_buffer_inflate_[toAddress(x, y, z)] == 0)
          continue;

        Eigen::Vector3d pos;
        indexToPos(Eigen::Vector3i(x, y, z), pos);
        if (pos(2) > mp_.visualization_truncate_height_)
          continue;

        pt.x = pos(0);
        pt.y = pos(1);
        pt.z = pos(2);
        cloud.push_back(pt);
      }

  cloud.width = cloud.points.size();
  cloud.height = 1;
  cloud.is_dense = true;
  cloud.header.frame_id = mp_.frame_id_;
  sensor_msgs::PointCloud2 cloud_msg;

  pcl::toROSMsg(cloud, cloud_msg);
  map_inf_pub_.publish(cloud_msg);

  // ROS_INFO("pub map");
}

void GridMap::publishUnknown()
{
  pcl::PointXYZ pt;
  pcl::PointCloud<pcl::PointXYZ> cloud;

  Eigen::Vector3i min_cut = md_.local_bound_min_;
  Eigen::Vector3i max_cut = md_.local_bound_max_;

  boundIndex(max_cut);
  boundIndex(min_cut);

  for (int x = min_cut(0); x <= max_cut(0); ++x)
    for (int y = min_cut(1); y <= max_cut(1); ++y)
      for (int z = min_cut(2); z <= max_cut(2); ++z)
      {

        if (md_.occupancy_buffer_[toAddress(x, y, z)] < mp_.clamp_min_log_ - 1e-3)
        {
          Eigen::Vector3d pos;
          indexToPos(Eigen::Vector3i(x, y, z), pos);
          if (pos(2) > mp_.visualization_truncate_height_)
            continue;

          pt.x = pos(0);
          pt.y = pos(1);
          pt.z = pos(2);
          cloud.push_back(pt);
        }
      }

  cloud.width = cloud.points.size();
  cloud.height = 1;
  cloud.is_dense = true;
  cloud.header.frame_id = mp_.frame_id_;

  sensor_msgs::PointCloud2 cloud_msg;
  pcl::toROSMsg(cloud, cloud_msg);
  unknown_pub_.publish(cloud_msg);
}

bool GridMap::odomValid() { return md_.has_odom_; }

bool GridMap::hasDepthObservation() { return md_.has_first_depth_; }

Eigen::Vector3d GridMap::getOrigin() { return mp_.map_origin_; }

// int GridMap::getVoxelNum() {
//   return mp_.map_voxel_num_[0] * mp_.map_voxel_num_[1] * mp_.map_voxel_num_[2];
// }

void GridMap::getRegion(Eigen::Vector3d &ori, Eigen::Vector3d &size)
{
  ori = mp_.map_origin_, size = mp_.map_size_;
}

void GridMap::depthOdomCallback(const sensor_msgs::ImageConstPtr &img,
                                const nav_msgs::OdometryConstPtr &odom)
{
  /* get pose */
  Eigen::Quaterniond body_q = Eigen::Quaterniond(odom->pose.pose.orientation.w,
                                                 odom->pose.pose.orientation.x,
                                                 odom->pose.pose.orientation.y,
                                                 odom->pose.pose.orientation.z);    
  Eigen::Matrix3d body_r_m = body_q.toRotationMatrix();   
  Eigen::Matrix4d body2world;
  body2world.block<3, 3>(0, 0) = body_r_m;
  body2world(0, 3) = odom->pose.pose.position.x;
  body2world(1, 3) = odom->pose.pose.position.y;
  body2world(2, 3) = odom->pose.pose.position.z;
  body2world(3, 3) = 1.0;
  
  Eigen::Matrix4d cam_T = body2world * md_.cam2body_;
  md_.camera_pos_(0) = cam_T(0, 3);
  md_.camera_pos_(1) = cam_T(1, 3);
  md_.camera_pos_(2) = cam_T(2, 3);
  md_.camera_q_ = Eigen::Quaterniond(cam_T.block<3, 3>(0, 0));

  /* get depth image */
  cv_bridge::CvImagePtr cv_ptr;
  cv_ptr = cv_bridge::toCvCopy(img, img->encoding);
  if (img->encoding == sensor_msgs::image_encodings::TYPE_32FC1)
  {
    (cv_ptr->image).convertTo(cv_ptr->image, CV_16UC1, mp_.k_depth_scaling_factor_);
  }
  cv_ptr->image.copyTo(md_.depth_image_);

  md_.occ_need_update_ = true;
}

// GridMap