// SPDX-License-Identifier: BSD-2-Clause

#include <ctime>
#include <mutex>
#include <atomic>
#include <memory>
#include <iomanip>
#include <iostream>
#include <unordered_map>
#include <boost/format.hpp>
#include <boost/thread.hpp>
#include <boost/filesystem.hpp>
#include <boost/algorithm/string.hpp>
#include <Eigen/Dense>
#include <pcl/io/pcd_io.h>

#include <ros/ros.h>
#include <geodesy/utm.h>
#include <geodesy/wgs84.h>
#include <pcl_ros/point_cloud.h>
#include <message_filters/subscriber.h>
#include <message_filters/time_synchronizer.h>
#include <message_filters/sync_policies/approximate_time.h>
#include <tf_conversions/tf_eigen.h>
#include <tf/transform_listener.h>

#include <std_msgs/Time.h>
#include <nav_msgs/Odometry.h>
#include <nmea_msgs/Sentence.h>
#include <sensor_msgs/Imu.h>
#include <sensor_msgs/NavSatFix.h>
#include <sensor_msgs/PointCloud2.h>
#include <geographic_msgs/GeoPointStamped.h>
#include <visualization_msgs/MarkerArray.h>
#include <hdl_graph_slam/FloorCoeffs.h>

#include <hdl_graph_slam/SaveMap.h>
#include <hdl_graph_slam/LoadGraph.h>
#include <hdl_graph_slam/DumpGraph.h>

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

#include <hdl_graph_slam/ros_utils.hpp>
#include <hdl_graph_slam/ros_time_hash.hpp>

#include <hdl_graph_slam/graph_slam.hpp>
#include <hdl_graph_slam/keyframe.hpp>
#include <hdl_graph_slam/keyframe_updater.hpp>
#include <hdl_graph_slam/loop_detector.hpp>
#include <hdl_graph_slam/information_matrix_calculator.hpp>
#include <hdl_graph_slam/map_cloud_generator.hpp>
#include <hdl_graph_slam/nmea_sentence_parser.hpp>

#include <g2o/types/slam3d/edge_se3.h>
#include <g2o/types/slam3d/vertex_se3.h>
#include <g2o/edge_se3_plane.hpp>
#include <g2o/edge_se3_priorxy.hpp>
#include <g2o/edge_se3_priorxyz.hpp>
#include <g2o/edge_se3_priorvec.hpp>
#include <g2o/edge_se3_priorquat.hpp>

/*
后端的作用(GPT):
1. 构建位资图:
  后端根据前端生成的关键帧位姿和相对位姿变换,构建位姿图(Pose Graph),每个关键帧对应图中的一个节点,节点之间的相对运动构成边。

2. 闭环检测
  后端通过检测闭环(即机器人经过某个已经访问过的地方)来减少累积误差。闭环检测通过计算当前帧和之前帧之间的相似度来发现闭环位置,并将其加入位姿图中,形成闭环约束边。

3. 图优化
  利用优化算法(如g2o或Ceres)对整个位姿图进行优化。目标是调整位姿图中所有关键帧的位姿,
  使之符合所有的约束(包括相邻关键帧的相对位姿约束以及闭环检测生成的约束),从而减少传感器噪声和漂移,生成全局一致的地图

 */

namespace hdl_graph_slam {

class HdlGraphSlamNodelet : public nodelet::Nodelet {
public:
  typedef pcl::PointXYZI PointT;
  typedef message_filters::sync_policies::ApproximateTime<nav_msgs::Odometry, sensor_msgs::PointCloud2> ApproxSyncPolicy;

  HdlGraphSlamNodelet() {}
  virtual ~HdlGraphSlamNodelet() {}

  virtual void onInit() {  // 初始化函数
    // 申请节点句柄
    nh = getNodeHandle();
    mt_nh = getMTNodeHandle();
    private_nh = getPrivateNodeHandle();

    // init parameters   初始化参数
    published_odom_topic = private_nh.param<std::string>("published_odom_topic", "/odom");
    map_frame_id = private_nh.param<std::string>("map_frame_id", "map");
    odom_frame_id = private_nh.param<std::string>("odom_frame_id", "odom");
    map_cloud_resolution = private_nh.param<double>("map_cloud_resolution", 0.05);
    trans_odom2map.setIdentity();

    max_keyframes_per_update = private_nh.param<int>("max_keyframes_per_update", 10);

    // 一些指针的初始化
    anchor_node = nullptr;  // 普通指针, 直接给 nullptr
    anchor_edge = nullptr;
    floor_plane_node = nullptr;
    graph_slam.reset(new GraphSLAM(private_nh.param<std::string>("g2o_solver_type", "lm_var")));  // 智能指针,用reset重新初始化,让它重新指向新分配的对象
    keyframe_updater.reset(new KeyframeUpdater(private_nh));
    loop_detector.reset(new LoopDetector(private_nh));
    map_cloud_generator.reset(new MapCloudGenerator());
    inf_calclator.reset(new InformationMatrixCalculator(private_nh));
    nmea_parser.reset(new NmeaSentenceParser());

    // 从参数服务器中读取参数
    // 第一个参数是参数服务器中的变量名,第二个参数是如果参数服务器中没有这个变量的默认取值
    gps_time_offset = private_nh.param<double>("gps_time_offset", 0.0);            // GPS时间戳与SLAM系统内部时间戳之间的偏移量,用于同步GPS数据和SLAM系统的时间
    gps_edge_stddev_xy = private_nh.param<double>("gps_edge_stddev_xy", 10000.0);  // GPS测量在XY平面(水平方向)的标准差,用于量化GPS测量的不确定性,通常用于优化算法中权重的计算
    gps_edge_stddev_z = private_nh.param<double>("gps_edge_stddev_z", 10.0);       // GPS测量在Z轴(垂直方向)的标准差,用于量化GPS测量的不确定性
    floor_edge_stddev = private_nh.param<double>("floor_edge_stddev", 10.0);       // 与地面相关的测量或估计的标准差,例如在检测到地面时用于优化算法

    imu_time_offset = private_nh.param<double>("imu_time_offset", 0.0);                            // IMU(惯性测量单元)时间戳与SLAM系统内部时间戳之间的偏移量,用于同步IMU数据和SLAM系统的时间
    enable_imu_orientation = private_nh.param<bool>("enable_imu_orientation", false);              // 指示是否启用IMU的方位数据
    enable_imu_acceleration = private_nh.param<bool>("enable_imu_acceleration", false);            // 指示是否启用IMU的加速度数据
    imu_orientation_edge_stddev = private_nh.param<double>("imu_orientation_edge_stddev", 0.1);    // IMU方位测量的标准差,用于量化IMU方位测量的不确定性
    imu_acceleration_edge_stddev = private_nh.param<double>("imu_acceleration_edge_stddev", 3.0);  // IMU加速度测量的标准差,用于量化IMU加速度测量的不确定性

    points_topic = private_nh.param<std::string>("points_topic", "/velodyne_points");  // 点云数据的ROS话题名称,SLAM系统将订阅这个话题来获取点云数据
    // 这个话题的info是:
    /* 
    (slam_env) mj@mj-Lenovo-Legion-R9000P2021H:~$ rostopic info /velodyne_points
    Type: sensor_msgs/PointCloud2

    Publishers: 
    * /player (http://mj-Lenovo-Legion-R9000P2021H:35735/)         
    这个发布方是 rosbag 呀,是没处理过的点云信息,后端怎么会订阅没处理过的点云数据?   
    也不是用来订阅的,这个文件是用这个话题来发布的,但是发布的内容又只有时间戳和frameid? 为什么这么奇怪?
    结合 read_until_pub 的解释来看,似乎是用来指示操作应该持续到何时,所以这个应该不用管

    Subscribers: 
    * /velodyne_nodelet_manager (http://mj-Lenovo-Legion-R9000P2021H:39235/)
    * /rviz_1729770290778318983 (http://mj-Lenovo-Legion-R9000P2021H:34629/)
    */



    // subscribers
    // 话题通信的接受方
    // 这个launch中的话题名是: odom
    odom_sub.reset(new message_filters::Subscriber<nav_msgs::Odometry>(mt_nh, published_odom_topic, 256));      // 创建一个新的message_filters::Subscriber对象,用于订阅nav_msgs::Odometry类型的消息
    cloud_sub.reset(new message_filters::Subscriber<sensor_msgs::PointCloud2>(mt_nh, "/filtered_points", 32));  // 创建一个新的message_filters::Subscriber对象,用于订阅sensor_msgs::PointCloud2类型的消息。
    // ApproxSyncPolicy 是一个同步策略,允许在一定时间窗口内的消息进行同步,这里设置的窗口大小为32
    sync.reset(new message_filters::Synchronizer<ApproxSyncPolicy>(ApproxSyncPolicy(32), *odom_sub, *cloud_sub));  // 创建一个新的message_filters::Synchronizer对象,用于同步上面创建的两个订阅者(odom_sub和cloud_sub)的消息。
    sync->registerCallback(boost::bind(&HdlGraphSlamNodelet::cloud_callback, this, _1, _2));                       // 为同步器注册了一个回调函数,当同步条件满足时,会调用HdlGraphSlamNodelet类的cloud_callback成员函数
    // 调用 cloud_callback 函数的条件:
    //    消息同步:odom_sub 和 cloud_sub 两个订阅者各自收到消息后,sync 会根据 ApproximateTime 策略尝试同步它们的时间戳。
    //    时间戳接近:两个消息的时间戳不需要完全一致,只要在一定范围内(由 ApproximateTime 策略自动决定),就会被认为是同步的。
    //    队列不为空:odom_sub 和 cloud_sub 的消息队列必须都各自至少包含一条消息,以便同步器能进行匹配。
    //    队列中找到匹配的消息对:当订阅者的队列中找到符合同步条件的消息对(即时间戳接近的 Odometry 和 PointCloud2 消息),就会触发 cloud_callback。

    imu_sub = nh.subscribe("/gpsimu_driver/imu_data", 1024, &HdlGraphSlamNodelet::imu_callback, this);                   // 接受并处理 imu 数据(小车没有imu,所以这个回调函数可以不看)
    floor_sub = nh.subscribe("/floor_detection/floor_coeffs", 1024, &HdlGraphSlamNodelet::floor_coeffs_callback, this);  // 接受并处理 地板检测 数据(小车没有地板检测,所以这个回调函数可以不看)

    if(private_nh.param<bool>("enable_gps", true)) {                                                                // !!!因为我们的小车没有GPS,所以这几个 callback 函数可以不看 !!!
      gps_sub = mt_nh.subscribe("/gps/geopoint", 1024, &HdlGraphSlamNodelet::gps_callback, this);                   // 接受并处理 GPS 数据
      nmea_sub = mt_nh.subscribe("/gpsimu_driver/nmea_sentence", 1024, &HdlGraphSlamNodelet::nmea_callback, this);  //
      navsat_sub = mt_nh.subscribe("/gps/navsat", 1024, &HdlGraphSlamNodelet::navsat_callback, this);               //
    }

    // publishers
    // 话题通信的发布方
    markers_pub = mt_nh.advertise<visualization_msgs::MarkerArray>("/hdl_graph_slam/markers", 16);      // 发布可视化数据(具体是什么还不知道)   这种类型的消息通常用于在RViz等可视化工具中显示标记(markers)
    odom2map_pub = mt_nh.advertise<geometry_msgs::TransformStamped>("/hdl_graph_slam/odom2pub", 16);    // 发布从里程计到地图的变换
    map_points_pub = mt_nh.advertise<sensor_msgs::PointCloud2>("/hdl_graph_slam/map_points", 1, true);  // 这个话题会保留以前扫描过的点云!! 扫过的地方的点云就是
    read_until_pub = mt_nh.advertise<std_msgs::Header>("/hdl_graph_slam/read_until", 32);               // 这种类型的消息通常用于指示读取操作应该持续到何时

    // 服务通信:一个节点(客户端)发送一个服务请求给服务服务器,服务服务器处理这个请求并返回一个响应。这是一种同步的、请求/响应模式的通信方式。
    load_service_server = mt_nh.advertiseService("/hdl_graph_slam/load", &HdlGraphSlamNodelet::load_service, this);  // 加载系统数据或状态 创建了一个服务服务器load_service_server,用于处理名为"/hdl_graph_slam/load"的服务请求,服务的处理函数是HdlGraphSlamNodelet类的load_service成员函数
    dump_service_server = mt_nh.advertiseService("/hdl_graph_slam/dump", &HdlGraphSlamNodelet::dump_service, this);  // 转储系统当前的状态或数据 创建了一个服务服务器dump_service_server,用于处理名为"/hdl_graph_slam/dump"的服务请求,服务的处理函数是HdlGraphSlamNodelet类的dump_service成员函数
    save_map_service_server = mt_nh.advertiseService("/hdl_graph_slam/save_map", &HdlGraphSlamNodelet::save_map_service, this);  // 保存地图

    graph_updated = false;                                                                           // 用于跟踪图优化是否已经更新
    double graph_update_interval = private_nh.param<double>("graph_update_interval", 3.0);           // 表示图优化更新的时间间隔,默认值为3.0秒    interval:间隔
    double map_cloud_update_interval = private_nh.param<double>("map_cloud_update_interval", 10.0);  // 地图点云更新的时间间隔,默认值为10.0秒

    // 主要是这个回调函数!!!
    optimization_timer = mt_nh.createWallTimer(ros::WallDuration(graph_update_interval), &HdlGraphSlamNodelet::optimization_timer_callback, this);  // 创建一个定时器,定时器会在每个 graph_update_interval 秒后触发,然后调用 optimization_timer_callback 函数

    map_publish_timer = mt_nh.createWallTimer(ros::WallDuration(map_cloud_update_interval), &HdlGraphSlamNodelet::map_points_publish_timer_callback, this);  // 创建一个定时器,定时器会在每个 map_cloud_update_interval 秒后触发,然后调用 map_points_publish_timer_callback 函数

    // 函数入口都是 callback 函数,也即三个: cloud_callback 、 optimization_timer_callback 、 map_points_publish_timer_callback
  }

private:
  /**
   * @brief received point clouds are pushed to #keyframe_queue
   * @param odom_msg  // 前端的激光里程计数据
   * @param cloud_msg // 前端滤波后的点云数据
   */
  void cloud_callback(const nav_msgs::OdometryConstPtr& odom_msg, const sensor_msgs::PointCloud2::ConstPtr& cloud_msg) {
    // 接收并处理激光里程计数据和滤波后的点云数据,并判断是否生成关键帧,然后将新生成的关键帧推送到 keyframe_queue 队列中
    // 处理点云信息主要是:格式转换、设置frame_id
    const ros::Time& stamp = cloud_msg->header.stamp;  // 获取点云数据的时间戳
    Eigen::Isometry3d odom = odom2isometry(odom_msg);  // 将里程计消息转换为Eigen库中的Isometry3d类型,它表示一个3D刚体变换(包括旋转和平移)

    pcl::PointCloud<PointT>::Ptr cloud(new pcl::PointCloud<PointT>());  // 创建一个点云对象的指针
    pcl::fromROSMsg(*cloud_msg, *cloud);                                // 将ROS的点云消息 cloud_msg 转换为PCL(Point Cloud Library)的点云格式
    if(base_frame_id.empty()) {
      base_frame_id = cloud_msg->header.frame_id;  // 将当前点云消息的坐标系 frame_id 设置为基础坐标系      base_frame_id 是整个系统中关键帧数据的参考坐标系
    }

    // 更新关键帧判断
    if(!keyframe_updater->update(odom)) {                      // 根据当前的里程计信息 odom 判断是否需要生成新的关键帧
      std::lock_guard<std::mutex> lock(keyframe_queue_mutex);  // 这行代码的作用是确保线程安全,防止多个线程同时访问或修改 keyframe_queue 队列
      // std::lock_guard<std::mutex> 是 C++ 标准库中的一个类,用来简化互斥锁(std::mutex)的使用。它的作用是在其作用域内自动对给定的互斥量加锁,并在作用域结束时自动解锁
      // keyframe_queue_mutex 是一个互斥量(std::mutex),用于保护关键帧队列 keyframe_queue。当多个线程试图同时访问或修改 keyframe_queue 时,互斥量确保只有一个线程能够访问该资源,避免数据竞争和并发问题
      // 工作原理:当这行代码执行时,lock_guard 会锁住 keyframe_queue_mutex,使其他线程无法访问与之关联的资源(即 keyframe_queue) 当程序执行离开该作用域时,lock_guard 会自动释放锁,无需显式调用 unlock() 函数
      // 为什么需要锁?   在多线程环境中(比如 ROS 回调函数可能被不同线程调用),如果多个线程同时读取或写入 keyframe_queue,可能导致数据不一致或崩溃。使用互斥量确保对 keyframe_queue 的操作是原子性的,即一个线程在操作时,其他线程必须等待

      if(keyframe_queue.empty()) {  // 如果关键帧队列是空的,发布一个 ROS 消息通知系统继续读取点云数据,直到指定时间点(stamp + ros::Duration(10, 0))。这段代码确保在没有关键帧生成时,系统能够继续读取点云数据
        std_msgs::Header read_until;
        read_until.stamp = stamp + ros::Duration(10, 0);
        read_until.frame_id = points_topic;
        read_until_pub.publish(read_until);
        read_until.frame_id = "/filtered_points";
        read_until_pub.publish(read_until);
      }

      return;
    }

    double accum_d = keyframe_updater->get_accum_distance();  // 获取累计的运动距离,用于判断关键帧生成的条件
    KeyFrame::Ptr keyframe(new KeyFrame(stamp, odom, accum_d, cloud));  // 创建一个新的关键帧 keyframe,其中包含当前的时间戳 stamp,里程计信息 odom,累计距离 accum_d,以及处理后的点云数据 cloud(这里的处理就是做了个消息类型的转换)

    std::lock_guard<std::mutex> lock(keyframe_queue_mutex);  // 使用 std::lock_guard 加锁,确保对关键帧队列 keyframe_queue 的操作是线程安全的
    keyframe_queue.push_back(keyframe);                      // 将新生成的关键帧 keyframe 推入 keyframe_queue 队列,供后续的处理使用
  }

  /**
   * @brief this method adds all the keyframes in #keyframe_queue to the pose graph (odometry edges)    此方法将#keyframe_queue中的所有关键帧添加到姿势图中(里程计边)
   * @return if true, at least one keyframe was added to the pose graph   如果为真,则至少有一个关键帧被添加到姿势图中
   * 将队列中的关键帧添加到位姿图中,并返回是否至少有一个关键帧被添加
   *
   * pose graph  是什么?
   */
  bool flush_keyframe_queue() {
    std::lock_guard<std::mutex> lock(keyframe_queue_mutex);  // 多线程锁

    if(keyframe_queue.empty()) {  // 没有关键帧就返回
      return false;
    }

    trans_odom2map_mutex.lock();
    Eigen::Isometry3d odom2map(trans_odom2map.cast<double>());
    trans_odom2map_mutex.unlock();

    std::cout << "flush_keyframe_queue - keyframes len:" << keyframes.size() << std::endl;
    int num_processed = 0;
    for(int i = 0; i < std::min<int>(keyframe_queue.size(), max_keyframes_per_update); i++) {
      num_processed = i;

      const auto& keyframe = keyframe_queue[i];
      // new_keyframes will be tested later for loop closure
      new_keyframes.push_back(keyframe);

      // add pose node
      Eigen::Isometry3d odom = odom2map * keyframe->odom;
      keyframe->node = graph_slam->add_se3_node(odom);
      keyframe_hash[keyframe->stamp] = keyframe;

      // fix the first node
      if(keyframes.empty() && new_keyframes.size() == 1) {
        if(private_nh.param<bool>("fix_first_node", false)) {
          Eigen::MatrixXd inf = Eigen::MatrixXd::Identity(6, 6);
          std::stringstream sst(private_nh.param<std::string>("fix_first_node_stddev", "1 1 1 1 1 1"));
          for(int i = 0; i < 6; i++) {
            double stddev = 1.0;
            sst >> stddev;
            inf(i, i) = 1.0 / stddev;
          }

          anchor_node = graph_slam->add_se3_node(Eigen::Isometry3d::Identity());
          anchor_node->setFixed(true);
          anchor_edge = graph_slam->add_se3_edge(anchor_node, keyframe->node, Eigen::Isometry3d::Identity(), inf);
        }
      }

      if(i == 0 && keyframes.empty()) {
        continue;
      }

      // add edge between consecutive keyframes
      const auto& prev_keyframe = i == 0 ? keyframes.back() : keyframe_queue[i - 1];

      Eigen::Isometry3d relative_pose = keyframe->odom.inverse() * prev_keyframe->odom;
      Eigen::MatrixXd information = inf_calclator->calc_information_matrix(keyframe->cloud, prev_keyframe->cloud, relative_pose);
      auto edge = graph_slam->add_se3_edge(keyframe->node, prev_keyframe->node, relative_pose, information);
      graph_slam->add_robust_kernel(edge, private_nh.param<std::string>("odometry_edge_robust_kernel", "NONE"), private_nh.param<double>("odometry_edge_robust_kernel_size", 1.0));
    }

    std_msgs::Header read_until;
    read_until.stamp = keyframe_queue[num_processed]->stamp + ros::Duration(10, 0);
    read_until.frame_id = points_topic;
    read_until_pub.publish(read_until);
    read_until.frame_id = "/filtered_points";
    read_until_pub.publish(read_until);

    keyframe_queue.erase(keyframe_queue.begin(), keyframe_queue.begin() + num_processed + 1);
    return true;
  }

  void nmea_callback(const nmea_msgs::SentenceConstPtr& nmea_msg) {
    GPRMC grmc = nmea_parser->parse(nmea_msg->sentence);

    if(grmc.status != 'A') {
      return;
    }

    geographic_msgs::GeoPointStampedPtr gps_msg(new geographic_msgs::GeoPointStamped());
    gps_msg->header = nmea_msg->header;
    gps_msg->position.latitude = grmc.latitude;
    gps_msg->position.longitude = grmc.longitude;
    gps_msg->position.altitude = NAN;

    gps_callback(gps_msg);
  }

  void navsat_callback(const sensor_msgs::NavSatFixConstPtr& navsat_msg) {
    geographic_msgs::GeoPointStampedPtr gps_msg(new geographic_msgs::GeoPointStamped());
    gps_msg->header = navsat_msg->header;
    gps_msg->position.latitude = navsat_msg->latitude;
    gps_msg->position.longitude = navsat_msg->longitude;
    gps_msg->position.altitude = navsat_msg->altitude;
    gps_callback(gps_msg);
  }

  /**
   * @brief received gps data is added to #gps_queue
   * @param gps_msg
   */
  void gps_callback(const geographic_msgs::GeoPointStampedPtr& gps_msg) {
    std::lock_guard<std::mutex> lock(gps_queue_mutex);
    gps_msg->header.stamp += ros::Duration(gps_time_offset);
    gps_queue.push_back(gps_msg);
  }

  /**
   * @brief
   * @return
   */
  bool flush_gps_queue() {
    std::lock_guard<std::mutex> lock(gps_queue_mutex);

    if(keyframes.empty() || gps_queue.empty()) {
      return false;
    }

    bool updated = false;
    auto gps_cursor = gps_queue.begin();

    for(auto& keyframe : keyframes) {
      if(keyframe->stamp > gps_queue.back()->header.stamp) {
        break;
      }

      if(keyframe->stamp < (*gps_cursor)->header.stamp || keyframe->utm_coord) {
        continue;
      }

      // find the gps data which is closest to the keyframe
      auto closest_gps = gps_cursor;
      for(auto gps = gps_cursor; gps != gps_queue.end(); gps++) {
        auto dt = ((*closest_gps)->header.stamp - keyframe->stamp).toSec();
        auto dt2 = ((*gps)->header.stamp - keyframe->stamp).toSec();
        if(std::abs(dt) < std::abs(dt2)) {
          break;
        }

        closest_gps = gps;
      }

      // if the time residual between the gps and keyframe is too large, skip it
      gps_cursor = closest_gps;
      if(0.2 < std::abs(((*closest_gps)->header.stamp - keyframe->stamp).toSec())) {
        continue;
      }

      // convert (latitude, longitude, altitude) -> (easting, northing, altitude) in UTM coordinate
      geodesy::UTMPoint utm;
      geodesy::fromMsg((*closest_gps)->position, utm);
      Eigen::Vector3d xyz(utm.easting, utm.northing, utm.altitude);

      // the first gps data position will be the origin of the map
      if(!zero_utm) {
        zero_utm = xyz;
      }
      xyz -= (*zero_utm);

      keyframe->utm_coord = xyz;

      g2o::OptimizableGraph::Edge* edge;
      if(std::isnan(xyz.z())) {
        Eigen::Matrix2d information_matrix = Eigen::Matrix2d::Identity() / gps_edge_stddev_xy;
        edge = graph_slam->add_se3_prior_xy_edge(keyframe->node, xyz.head<2>(), information_matrix);
      } else {
        Eigen::Matrix3d information_matrix = Eigen::Matrix3d::Identity();
        information_matrix.block<2, 2>(0, 0) /= gps_edge_stddev_xy;
        information_matrix(2, 2) /= gps_edge_stddev_z;
        edge = graph_slam->add_se3_prior_xyz_edge(keyframe->node, xyz, information_matrix);
      }
      graph_slam->add_robust_kernel(edge, private_nh.param<std::string>("gps_edge_robust_kernel", "NONE"), private_nh.param<double>("gps_edge_robust_kernel_size", 1.0));

      updated = true;
    }

    auto remove_loc = std::upper_bound(gps_queue.begin(), gps_queue.end(), keyframes.back()->stamp, [=](const ros::Time& stamp, const geographic_msgs::GeoPointStampedConstPtr& geopoint) { return stamp < geopoint->header.stamp; });
    gps_queue.erase(gps_queue.begin(), remove_loc);
    return updated;
  }

  void imu_callback(const sensor_msgs::ImuPtr& imu_msg) {
    if(!enable_imu_orientation && !enable_imu_acceleration) {
      return;
    }

    std::lock_guard<std::mutex> lock(imu_queue_mutex);
    imu_msg->header.stamp += ros::Duration(imu_time_offset);
    imu_queue.push_back(imu_msg);
  }

  bool flush_imu_queue() {
    std::lock_guard<std::mutex> lock(imu_queue_mutex);
    if(keyframes.empty() || imu_queue.empty() || base_frame_id.empty()) {
      return false;
    }

    bool updated = false;
    auto imu_cursor = imu_queue.begin();

    for(auto& keyframe : keyframes) {
      if(keyframe->stamp > imu_queue.back()->header.stamp) {
        break;
      }

      if(keyframe->stamp < (*imu_cursor)->header.stamp || keyframe->acceleration) {
        continue;
      }

      // find imu data which is closest to the keyframe
      auto closest_imu = imu_cursor;
      for(auto imu = imu_cursor; imu != imu_queue.end(); imu++) {
        auto dt = ((*closest_imu)->header.stamp - keyframe->stamp).toSec();
        auto dt2 = ((*imu)->header.stamp - keyframe->stamp).toSec();
        if(std::abs(dt) < std::abs(dt2)) {
          break;
        }

        closest_imu = imu;
      }

      imu_cursor = closest_imu;
      if(0.2 < std::abs(((*closest_imu)->header.stamp - keyframe->stamp).toSec())) {
        continue;
      }

      const auto& imu_ori = (*closest_imu)->orientation;
      const auto& imu_acc = (*closest_imu)->linear_acceleration;

      geometry_msgs::Vector3Stamped acc_imu;
      geometry_msgs::Vector3Stamped acc_base;
      geometry_msgs::QuaternionStamped quat_imu;
      geometry_msgs::QuaternionStamped quat_base;

      quat_imu.header.frame_id = acc_imu.header.frame_id = (*closest_imu)->header.frame_id;
      quat_imu.header.stamp = acc_imu.header.stamp = ros::Time(0);
      acc_imu.vector = (*closest_imu)->linear_acceleration;
      quat_imu.quaternion = (*closest_imu)->orientation;

      try {
        tf_listener.transformVector(base_frame_id, acc_imu, acc_base);
        tf_listener.transformQuaternion(base_frame_id, quat_imu, quat_base);
      } catch(std::exception& e) {
        std::cerr << "failed to find transform!!" << std::endl;
        return false;
      }

      keyframe->acceleration = Eigen::Vector3d(acc_base.vector.x, acc_base.vector.y, acc_base.vector.z);
      keyframe->orientation = Eigen::Quaterniond(quat_base.quaternion.w, quat_base.quaternion.x, quat_base.quaternion.y, quat_base.quaternion.z);
      keyframe->orientation = keyframe->orientation;
      if(keyframe->orientation->w() < 0.0) {
        keyframe->orientation->coeffs() = -keyframe->orientation->coeffs();
      }

      if(enable_imu_orientation) {
        Eigen::MatrixXd info = Eigen::MatrixXd::Identity(3, 3) / imu_orientation_edge_stddev;
        auto edge = graph_slam->add_se3_prior_quat_edge(keyframe->node, *keyframe->orientation, info);
        graph_slam->add_robust_kernel(edge, private_nh.param<std::string>("imu_orientation_edge_robust_kernel", "NONE"), private_nh.param<double>("imu_orientation_edge_robust_kernel_size", 1.0));
      }

      if(enable_imu_acceleration) {
        Eigen::MatrixXd info = Eigen::MatrixXd::Identity(3, 3) / imu_acceleration_edge_stddev;
        g2o::OptimizableGraph::Edge* edge = graph_slam->add_se3_prior_vec_edge(keyframe->node, -Eigen::Vector3d::UnitZ(), *keyframe->acceleration, info);
        graph_slam->add_robust_kernel(edge, private_nh.param<std::string>("imu_acceleration_edge_robust_kernel", "NONE"), private_nh.param<double>("imu_acceleration_edge_robust_kernel_size", 1.0));
      }
      updated = true;
    }

    auto remove_loc = std::upper_bound(imu_queue.begin(), imu_queue.end(), keyframes.back()->stamp, [=](const ros::Time& stamp, const sensor_msgs::ImuConstPtr& imu) { return stamp < imu->header.stamp; });
    imu_queue.erase(imu_queue.begin(), remove_loc);

    return updated;
  }

  /**   不做地板检测,这个函数不用看
   * @brief received floor coefficients are added to #floor_coeffs_queue
   * @param floor_coeffs_msg
   */
  void floor_coeffs_callback(const hdl_graph_slam::FloorCoeffsConstPtr& floor_coeffs_msg) {
    if(floor_coeffs_msg->coeffs.empty()) {
      return;
    }

    std::lock_guard<std::mutex> lock(floor_coeffs_queue_mutex);
    floor_coeffs_queue.push_back(floor_coeffs_msg);
  }

  /**   不做地板检测,这个函数不用看
   * @brief this methods associates floor coefficients messages with registered keyframes, and then adds the associated coeffs to the pose graph
   * @brief 该方法将地板系数消息与注册的关键帧相关联,然后将关联的系数添加到姿势图中。
   * @return if true, at least one floor plane edge is added to the pose graph
   * @return 如果为真,则至少有一个地板平面边被添加到姿势图中
   */
  bool flush_floor_queue() {
    std::lock_guard<std::mutex> lock(floor_coeffs_queue_mutex);

    if(keyframes.empty()) {
      return false;
    }

    const auto& latest_keyframe_stamp = keyframes.back()->stamp;

    bool updated = false;
    for(const auto& floor_coeffs : floor_coeffs_queue) {
      if(floor_coeffs->header.stamp > latest_keyframe_stamp) {
        break;
      }

      auto found = keyframe_hash.find(floor_coeffs->header.stamp);
      if(found == keyframe_hash.end()) {
        continue;
      }

      if(!floor_plane_node) {
        floor_plane_node = graph_slam->add_plane_node(Eigen::Vector4d(0.0, 0.0, 1.0, 0.0));
        floor_plane_node->setFixed(true);
      }

      const auto& keyframe = found->second;

      Eigen::Vector4d coeffs(floor_coeffs->coeffs[0], floor_coeffs->coeffs[1], floor_coeffs->coeffs[2], floor_coeffs->coeffs[3]);
      Eigen::Matrix3d information = Eigen::Matrix3d::Identity() * (1.0 / floor_edge_stddev);
      auto edge = graph_slam->add_se3_plane_edge(keyframe->node, floor_plane_node, coeffs, information);
      graph_slam->add_robust_kernel(edge, private_nh.param<std::string>("floor_edge_robust_kernel", "NONE"), private_nh.param<double>("floor_edge_robust_kernel_size", 1.0));

      keyframe->floor_coeffs = coeffs;

      updated = true;
    }

    auto remove_loc = std::upper_bound(floor_coeffs_queue.begin(), floor_coeffs_queue.end(), latest_keyframe_stamp, [=](const ros::Time& stamp, const hdl_graph_slam::FloorCoeffsConstPtr& coeffs) { return stamp < coeffs->header.stamp; });
    floor_coeffs_queue.erase(floor_coeffs_queue.begin(), remove_loc);

    return updated;
  }

  /**
   * @brief generate map point cloud and publish it
   * @param event
   */
  void map_points_publish_timer_callback(const ros::WallTimerEvent& event) {
    if(!map_points_pub.getNumSubscribers() || !graph_updated) {
      return;
    }

    std::vector<KeyFrameSnapshot::Ptr> snapshot;

    keyframes_snapshot_mutex.lock();
    snapshot = keyframes_snapshot;
    keyframes_snapshot_mutex.unlock();

    auto cloud = map_cloud_generator->generate(snapshot, map_cloud_resolution);
    if(!cloud) {
      return;
    }

    cloud->header.frame_id = map_frame_id;
    cloud->header.stamp = snapshot.back()->cloud->header.stamp;

    sensor_msgs::PointCloud2Ptr cloud_msg(new sensor_msgs::PointCloud2());
    pcl::toROSMsg(*cloud, *cloud_msg);

    map_points_pub.publish(cloud_msg);
  }

  /**
   * @brief this methods adds all the data in the queues to the pose graph, and then optimizes the pose graph
   * @brief 此方法将队列中的所有数据添加到姿势图中,然后优化姿势图
   * @param event   // 是ROS提供的定时器时间类,包含信息有:计时器出发的时间戳、上一次出发的时间戳、计时器的周期信息等
   */
  void optimization_timer_callback(const ros::WallTimerEvent& event) {
    std::lock_guard<std::mutex> lock(main_thread_mutex);  // 创建了一个锁,用于保护对主线程的访问,确保线程安全

    // add keyframes and floor coeffs in the queues to the pose graph  将队列中的关键帧和楼层系数添加到姿势图中
    bool keyframe_updated = flush_keyframe_queue();  // 调用 flush_keyframe_queue 函数,将关键帧队列中的数据添加到位姿图中,并返回是否有关键帧被更新

    if(!keyframe_updated) {  // 如果没有关键帧被更新,则执行大括号内的代码
      std_msgs::Header read_until;
      read_until.stamp = ros::Time::now() + ros::Duration(30, 0);  // 时间戳为什么要加30秒?
      read_until.frame_id = points_topic;
      read_until_pub.publish(read_until);
      read_until.frame_id = "/filtered_points";
      read_until_pub.publish(read_until);  // 不同的 frame_id 和 话题 发布两次, 是干嘛用的?
    }

    if(!keyframe_updated & !flush_floor_queue() & !flush_gps_queue() & !flush_imu_queue()) {
      // 检查地板、GPS、IMU等队列,分别调用相应的 flush 函数(例如 flush_floor_queue、flush_gps_queue、flush_imu_queue),如果这些数据没有更新,则函数直接返回,不进行后续操作。
      return;
    }

    // loop detection   闭环检测
    std::vector<Loop::Ptr> loops = loop_detector->detect(keyframes, new_keyframes, *graph_slam);
    // 使用 loop_detector->detect 检测关键帧之间的闭环。闭环检测用于检测机器人是否回到了之前的某个位置,这在位姿图优化中起到约束作用。  (这部分都是GPT解释)
    // keyframes: 包含所有历史关键帧的容器
    // new_keyframes: 当前生成的最新的关键帧
    // *graph_slam: 一个图优化的实例,用来存储闭环检测的结果,可能包括相对位资信息,图结构,以及优化过程中所需要的数据
    // 相对位资信息: new_keyframes 与 keyframes 中所有帧的相对位资
    // 图结构: 一种数学表达,用来描述机器人在环境中的位资及其与环境中特征(如地标和其他关键帧)的关系. 具体而言,包括下面几个基本要素:
    //    节点: 每个节点代表一个特定的位资或状态  例如slam中,一个节点可以表示机器人在某个点的位置信息,也可以表示环境中的特征点
    //    边: 表示节点之间的约束关系, 这些约束可以是基于传感器测量的数据(例如激光里程计)或者闭环检测结果.    边通常回代有权重,表示约束的可信度和不确定性
    //    优化目标: 通过最小化节点之间的约束误差,图结构可以通过图优化算法进行优化.  常见的方法包括最小二乘法, 梯度下降法等等.   优化的目标是使得地图和机器人轨迹尽可能一致, 提高整体的定位和建图精度
    //    闭环检测: 在闭环检测过程中, 图结构会被更新,以反映新的闭环约束.这意味着新的边被添加到图中, 链接当前节点和历史节点,改善整体地图的一致性
    //    数据关联: 通过匹配特征和关键帧, 确定节点之间的关联性,从而形成合理的地图和路径

    // 检测不到闭环的情况: 没有足够的关键帧、环境发生显著变化、特征匹配失败(发生错误或特征点数量不足)、定位精度不足、参数设置不当、移动机器人的运动太快、内存或计算资源不够

    // 如果检测到闭环,则为闭环关系添加一个新的边到位姿图中,计算相对位姿并生成信息矩阵,确保闭环边的优化权重。
    for(const auto& loop : loops) {  // 遍历闭环检测结果
      Eigen::Isometry3d relpose(loop->relative_pose.cast<double>());
      Eigen::MatrixXd information_matrix = inf_calclator->calc_information_matrix(loop->key1->cloud, loop->key2->cloud, relpose);
      auto edge = graph_slam->add_se3_edge(loop->key1->node, loop->key2->node, relpose, information_matrix);
      graph_slam->add_robust_kernel(edge, private_nh.param<std::string>("loop_closure_edge_robust_kernel", "NONE"), private_nh.param<double>("loop_closure_edge_robust_kernel_size", 1.0));
    }

    std::copy(new_keyframes.begin(), new_keyframes.end(), std::back_inserter(keyframes));  // 将新关键帧 new_keyframes 合并到已有的关键帧列表 keyframes 中
    new_keyframes.clear();                                                                 // 清空 new_keyframes

    // move the first node anchor position to the current estimate of the first node pose   将第一节点锚点位置移动到第一节点姿态的当前估计值
    // so the first node moves freely while trying to stay around the origin                因此,第一个节点在试图停留在原点附近的同时可以自由移动

    if(anchor_node && private_nh.param<bool>("fix_first_node_adaptive", true)) {  // launch文件中,fix_first_node_adaptive 设置为 true
      Eigen::Isometry3d anchor_target = static_cast<g2o::VertexSE3*>(anchor_edge->vertices()[1])->estimate();
      anchor_node->setEstimate(anchor_target);
      // 如果启用了自适应固定第一帧的功能(参数 "fix_first_node_adaptive"),则将第一个关键帧的锚点位置更新为当前估计的位置,以允许它自由移动但仍然保持在原点附近。
    }

    // optimize the pose graph
    int num_iterations = private_nh.param<int>("g2o_solver_num_iterations", 1024);  // launch文件中都是设置成512
    graph_slam->optimize(num_iterations); // 使用 g2o 优化器对位姿图进行优化,优化的迭代次数由参数 "g2o_solver_num_iterations" 控制

    // publish tf     发布位姿变换
    const auto& keyframe = keyframes.back();  // 获取最新的关键帧估计
    Eigen::Isometry3d trans = keyframe->node->estimate() * keyframe->odom.inverse();
    trans_odom2map_mutex.lock();
    trans_odom2map = trans.matrix().cast<float>();
    trans_odom2map_mutex.unlock();

    std::vector<KeyFrameSnapshot::Ptr> snapshot(keyframes.size());
    std::transform(keyframes.begin(), keyframes.end(), snapshot.begin(), [=](const KeyFrame::Ptr& k) { return std::make_shared<KeyFrameSnapshot>(k); });

    keyframes_snapshot_mutex.lock();
    keyframes_snapshot.swap(snapshot);
    keyframes_snapshot_mutex.unlock();
    graph_updated = true;
    

    if(odom2map_pub.getNumSubscribers()) {
      geometry_msgs::TransformStamped ts = matrix2transform(keyframe->stamp, trans.matrix().cast<float>(), map_frame_id, odom_frame_id);
      odom2map_pub.publish(ts);  // 发布odom2map_pub话题中
    }

    if(markers_pub.getNumSubscribers()) {
      auto markers = create_marker_array(ros::Time::now());
      markers_pub.publish(markers);
    }
  }

  /**
   * @brief create visualization marker
   * @param stamp
   * @return
   */
  visualization_msgs::MarkerArray create_marker_array(const ros::Time& stamp) const {
    visualization_msgs::MarkerArray markers;
    markers.markers.resize(4);

    // node markers
    visualization_msgs::Marker& traj_marker = markers.markers[0];
    traj_marker.header.frame_id = "map";
    traj_marker.header.stamp = stamp;
    traj_marker.ns = "nodes";
    traj_marker.id = 0;
    traj_marker.type = visualization_msgs::Marker::SPHERE_LIST;

    traj_marker.pose.orientation.w = 1.0;
    traj_marker.scale.x = traj_marker.scale.y = traj_marker.scale.z = 0.5;

    visualization_msgs::Marker& imu_marker = markers.markers[1];
    imu_marker.header = traj_marker.header;
    imu_marker.ns = "imu";
    imu_marker.id = 1;
    imu_marker.type = visualization_msgs::Marker::SPHERE_LIST;

    imu_marker.pose.orientation.w = 1.0;
    imu_marker.scale.x = imu_marker.scale.y = imu_marker.scale.z = 0.75;

    traj_marker.points.resize(keyframes.size());
    traj_marker.colors.resize(keyframes.size());
    for(int i = 0; i < keyframes.size(); i++) {
      Eigen::Vector3d pos = keyframes[i]->node->estimate().translation();
      traj_marker.points[i].x = pos.x();
      traj_marker.points[i].y = pos.y();
      traj_marker.points[i].z = pos.z();

      double p = static_cast<double>(i) / keyframes.size();
      traj_marker.colors[i].r = 1.0 - p;
      traj_marker.colors[i].g = p;
      traj_marker.colors[i].b = 0.0;
      traj_marker.colors[i].a = 1.0;

      if(keyframes[i]->acceleration) {
        Eigen::Vector3d pos = keyframes[i]->node->estimate().translation();
        geometry_msgs::Point point;
        point.x = pos.x();
        point.y = pos.y();
        point.z = pos.z();

        std_msgs::ColorRGBA color;
        color.r = 0.0;
        color.g = 0.0;
        color.b = 1.0;
        color.a = 0.1;

        imu_marker.points.push_back(point);
        imu_marker.colors.push_back(color);
      }
    }

    // edge markers
    visualization_msgs::Marker& edge_marker = markers.markers[2];
    edge_marker.header.frame_id = "map";
    edge_marker.header.stamp = stamp;
    edge_marker.ns = "edges";
    edge_marker.id = 2;
    edge_marker.type = visualization_msgs::Marker::LINE_LIST;

    edge_marker.pose.orientation.w = 1.0;
    edge_marker.scale.x = 0.05;

    edge_marker.points.resize(graph_slam->graph->edges().size() * 2);
    edge_marker.colors.resize(graph_slam->graph->edges().size() * 2);

    auto edge_itr = graph_slam->graph->edges().begin();
    for(int i = 0; edge_itr != graph_slam->graph->edges().end(); edge_itr++, i++) {
      g2o::HyperGraph::Edge* edge = *edge_itr;
      g2o::EdgeSE3* edge_se3 = dynamic_cast<g2o::EdgeSE3*>(edge);
      if(edge_se3) {
        g2o::VertexSE3* v1 = dynamic_cast<g2o::VertexSE3*>(edge_se3->vertices()[0]);
        g2o::VertexSE3* v2 = dynamic_cast<g2o::VertexSE3*>(edge_se3->vertices()[1]);
        Eigen::Vector3d pt1 = v1->estimate().translation();
        Eigen::Vector3d pt2 = v2->estimate().translation();

        edge_marker.points[i * 2].x = pt1.x();
        edge_marker.points[i * 2].y = pt1.y();
        edge_marker.points[i * 2].z = pt1.z();
        edge_marker.points[i * 2 + 1].x = pt2.x();
        edge_marker.points[i * 2 + 1].y = pt2.y();
        edge_marker.points[i * 2 + 1].z = pt2.z();

        double p1 = static_cast<double>(v1->id()) / graph_slam->graph->vertices().size();
        double p2 = static_cast<double>(v2->id()) / graph_slam->graph->vertices().size();
        edge_marker.colors[i * 2].r = 1.0 - p1;
        edge_marker.colors[i * 2].g = p1;
        edge_marker.colors[i * 2].a = 1.0;
        edge_marker.colors[i * 2 + 1].r = 1.0 - p2;
        edge_marker.colors[i * 2 + 1].g = p2;
        edge_marker.colors[i * 2 + 1].a = 1.0;

        if(std::abs(v1->id() - v2->id()) > 2) {
          edge_marker.points[i * 2].z += 0.5;
          edge_marker.points[i * 2 + 1].z += 0.5;
        }

        continue;
      }

      g2o::EdgeSE3Plane* edge_plane = dynamic_cast<g2o::EdgeSE3Plane*>(edge);
      if(edge_plane) {
        g2o::VertexSE3* v1 = dynamic_cast<g2o::VertexSE3*>(edge_plane->vertices()[0]);
        Eigen::Vector3d pt1 = v1->estimate().translation();
        Eigen::Vector3d pt2(pt1.x(), pt1.y(), 0.0);

        edge_marker.points[i * 2].x = pt1.x();
        edge_marker.points[i * 2].y = pt1.y();
        edge_marker.points[i * 2].z = pt1.z();
        edge_marker.points[i * 2 + 1].x = pt2.x();
        edge_marker.points[i * 2 + 1].y = pt2.y();
        edge_marker.points[i * 2 + 1].z = pt2.z();

        edge_marker.colors[i * 2].b = 1.0;
        edge_marker.colors[i * 2].a = 1.0;
        edge_marker.colors[i * 2 + 1].b = 1.0;
        edge_marker.colors[i * 2 + 1].a = 1.0;

        continue;
      }

      g2o::EdgeSE3PriorXY* edge_priori_xy = dynamic_cast<g2o::EdgeSE3PriorXY*>(edge);
      if(edge_priori_xy) {
        g2o::VertexSE3* v1 = dynamic_cast<g2o::VertexSE3*>(edge_priori_xy->vertices()[0]);
        Eigen::Vector3d pt1 = v1->estimate().translation();
        Eigen::Vector3d pt2 = Eigen::Vector3d::Zero();
        pt2.head<2>() = edge_priori_xy->measurement();

        edge_marker.points[i * 2].x = pt1.x();
        edge_marker.points[i * 2].y = pt1.y();
        edge_marker.points[i * 2].z = pt1.z() + 0.5;
        edge_marker.points[i * 2 + 1].x = pt2.x();
        edge_marker.points[i * 2 + 1].y = pt2.y();
        edge_marker.points[i * 2 + 1].z = pt2.z() + 0.5;

        edge_marker.colors[i * 2].r = 1.0;
        edge_marker.colors[i * 2].a = 1.0;
        edge_marker.colors[i * 2 + 1].r = 1.0;
        edge_marker.colors[i * 2 + 1].a = 1.0;

        continue;
      }

      g2o::EdgeSE3PriorXYZ* edge_priori_xyz = dynamic_cast<g2o::EdgeSE3PriorXYZ*>(edge);
      if(edge_priori_xyz) {
        g2o::VertexSE3* v1 = dynamic_cast<g2o::VertexSE3*>(edge_priori_xyz->vertices()[0]);
        Eigen::Vector3d pt1 = v1->estimate().translation();
        Eigen::Vector3d pt2 = edge_priori_xyz->measurement();

        edge_marker.points[i * 2].x = pt1.x();
        edge_marker.points[i * 2].y = pt1.y();
        edge_marker.points[i * 2].z = pt1.z() + 0.5;
        edge_marker.points[i * 2 + 1].x = pt2.x();
        edge_marker.points[i * 2 + 1].y = pt2.y();
        edge_marker.points[i * 2 + 1].z = pt2.z();

        edge_marker.colors[i * 2].r = 1.0;
        edge_marker.colors[i * 2].a = 1.0;
        edge_marker.colors[i * 2 + 1].r = 1.0;
        edge_marker.colors[i * 2 + 1].a = 1.0;

        continue;
      }
    }

    // sphere
    visualization_msgs::Marker& sphere_marker = markers.markers[3];
    sphere_marker.header.frame_id = "map";
    sphere_marker.header.stamp = stamp;
    sphere_marker.ns = "loop_close_radius";
    sphere_marker.id = 3;
    sphere_marker.type = visualization_msgs::Marker::SPHERE;

    if(!keyframes.empty()) {
      Eigen::Vector3d pos = keyframes.back()->node->estimate().translation();
      sphere_marker.pose.position.x = pos.x();
      sphere_marker.pose.position.y = pos.y();
      sphere_marker.pose.position.z = pos.z();
    }
    sphere_marker.pose.orientation.w = 1.0;
    sphere_marker.scale.x = sphere_marker.scale.y = sphere_marker.scale.z = loop_detector->get_distance_thresh() * 2.0;

    sphere_marker.color.r = 1.0;
    sphere_marker.color.a = 0.3;

    return markers;
  }

  /**
   * @brief load all data from a directory  从目录加载所有数据
   * @param req
   * @param res
   * @return
   */
  bool load_service(hdl_graph_slam::LoadGraphRequest& req, hdl_graph_slam::LoadGraphResponse& res) {
    std::lock_guard<std::mutex> lock(main_thread_mutex);

    std::string directory = req.path;

    std::cout << "loading data from:" << directory << std::endl;

    // Load graph.
    graph_slam->load(directory + "/graph.g2o");

    // Iterate over the items in this directory and count how many sub directories there are.
    // This will give an upper limit on how many keyframe indexes we can expect to find.
    boost::filesystem::directory_iterator begin(directory), end;
    int max_directory_count = std::count_if(begin, end, [](const boost::filesystem::directory_entry& d) {
      return boost::filesystem::is_directory(d.path());  // only return true if a direcotry
    });

    // Load keyframes by looping through key frame indexes that we expect to see.
    for(int i = 0; i < max_directory_count; i++) {
      std::stringstream sst;
      sst << boost::format("%s/%06d") % directory % i;
      std::string key_frame_directory = sst.str();

      // If key_frame_directory doesnt exist, then we have run out so lets stop looking.
      if(!boost::filesystem::is_directory(key_frame_directory)) {
        break;
      }

      KeyFrame::Ptr keyframe(new KeyFrame(key_frame_directory, graph_slam->graph.get()));
      keyframes.push_back(keyframe);
    }
    std::cout << "loaded " << keyframes.size() << " keyframes" << std::endl;

    // Load special nodes.
    std::ifstream ifs(directory + "/special_nodes.csv");
    if(!ifs) {
      return false;
    }
    while(!ifs.eof()) {
      std::string token;
      ifs >> token;
      if(token == "anchor_node") {
        int id = 0;
        ifs >> id;
        anchor_node = static_cast<g2o::VertexSE3*>(graph_slam->graph->vertex(id));
      } else if(token == "anchor_edge") {
        int id = 0;
        ifs >> id;
        // We have no way of directly pulling the edge based on the edge ID that we have just read in.
        // Fortunatly anchor edges are always attached to the anchor node so we can iterate over
        // the edges that listed against the node untill we find the one that matches our ID.
        if(anchor_node) {
          auto edges = anchor_node->edges();

          for(auto e : edges) {
            int edgeID = e->id();
            if(edgeID == id) {
              anchor_edge = static_cast<g2o::EdgeSE3*>(e);

              break;
            }
          }
        }
      } else if(token == "floor_node") {
        int id = 0;
        ifs >> id;
        floor_plane_node = static_cast<g2o::VertexPlane*>(graph_slam->graph->vertex(id));
      }
    }

    // check if we have any non null special nodes, if all are null then lets not bother.
    if(anchor_node->id() || anchor_edge->id() || floor_plane_node->id()) {
      std::cout << "loaded special nodes - ";

      // check exists before printing information about each special node
      if(anchor_node->id()) {
        std::cout << " anchor_node: " << anchor_node->id();
      }
      if(anchor_edge->id()) {
        std::cout << " anchor_edge: " << anchor_edge->id();
      }
      if(floor_plane_node->id()) {
        std::cout << " floor_node: " << floor_plane_node->id();
      }

      // finish with a new line
      std::cout << std::endl;
    }

    // Update our keyframe snapshot so we can publish a map update, trigger update with graph_updated = true.
    std::vector<KeyFrameSnapshot::Ptr> snapshot(keyframes.size());

    std::transform(keyframes.begin(), keyframes.end(), snapshot.begin(), [=](const KeyFrame::Ptr& k) { return std::make_shared<KeyFrameSnapshot>(k); });

    keyframes_snapshot_mutex.lock();
    keyframes_snapshot.swap(snapshot);
    keyframes_snapshot_mutex.unlock();
    graph_updated = true;

    res.success = true;

    std::cout << "snapshot updated" << std::endl << "loading successful" << std::endl;

    return true;
  }

  /**
   * @brief dump all data to the current directory    将所有数据转储到当前目录
   * @param req
   * @param res
   * @return
   */
  bool dump_service(hdl_graph_slam::DumpGraphRequest& req, hdl_graph_slam::DumpGraphResponse& res) {
    std::lock_guard<std::mutex> lock(main_thread_mutex);

    std::string directory = req.destination;

    if(directory.empty()) {
      std::array<char, 64> buffer;
      buffer.fill(0);
      time_t rawtime;
      time(&rawtime);
      const auto timeinfo = localtime(&rawtime);
      strftime(buffer.data(), sizeof(buffer), "%d-%m-%Y %H:%M:%S", timeinfo);
    }

    if(!boost::filesystem::is_directory(directory)) {
      boost::filesystem::create_directory(directory);
    }

    std::cout << "dumping data to:" << directory << std::endl;
    // save graph
    graph_slam->save(directory + "/graph.g2o");

    // save keyframes
    for(int i = 0; i < keyframes.size(); i++) {
      std::stringstream sst;
      sst << boost::format("%s/%06d") % directory % i;

      keyframes[i]->save(sst.str());
    }

    if(zero_utm) {
      std::ofstream zero_utm_ofs(directory + "/zero_utm");
      zero_utm_ofs << boost::format("%.6f %.6f %.6f") % zero_utm->x() % zero_utm->y() % zero_utm->z() << std::endl;
    }

    std::ofstream ofs(directory + "/special_nodes.csv");
    ofs << "anchor_node " << (anchor_node == nullptr ? -1 : anchor_node->id()) << std::endl;
    ofs << "anchor_edge " << (anchor_edge == nullptr ? -1 : anchor_edge->id()) << std::endl;
    ofs << "floor_node " << (floor_plane_node == nullptr ? -1 : floor_plane_node->id()) << std::endl;

    res.success = true;
    return true;
  }

  /**
   * @brief save map data as pcd
   * @param req
   * @param res
   * @return
   */
  bool save_map_service(hdl_graph_slam::SaveMapRequest& req, hdl_graph_slam::SaveMapResponse& res) {
    std::vector<KeyFrameSnapshot::Ptr> snapshot;

    keyframes_snapshot_mutex.lock();
    snapshot = keyframes_snapshot;
    keyframes_snapshot_mutex.unlock();

    auto cloud = map_cloud_generator->generate(snapshot, req.resolution);
    if(!cloud) {
      res.success = false;
      return true;
    }

    if(zero_utm && req.utm) {
      for(auto& pt : cloud->points) {
        pt.getVector3fMap() += (*zero_utm).cast<float>();
      }
    }

    cloud->header.frame_id = map_frame_id;
    cloud->header.stamp = snapshot.back()->cloud->header.stamp;

    if(zero_utm) {
      std::ofstream ofs(req.destination + ".utm");
      ofs << boost::format("%.6f %.6f %.6f") % zero_utm->x() % zero_utm->y() % zero_utm->z() << std::endl;
    }

    int ret = pcl::io::savePCDFileBinary(req.destination, *cloud);
    res.success = ret == 0;

    return true;
  }

private:  // 一些参数的定义
  // ROS
  ros::NodeHandle nh;          // 节点句柄
  ros::NodeHandle mt_nh;       //
  ros::NodeHandle private_nh;  // 私有节点句柄     申请多几个节点句柄主要是为了:命名空间隔离、参数隔离、代码组织、责任分离、调试和测试、灵活性和扩展性、并行处理(kimi解释)
  ros::WallTimer optimization_timer;
  ros::WallTimer map_publish_timer;

  std::unique_ptr<message_filters::Subscriber<nav_msgs::Odometry>> odom_sub;
  std::unique_ptr<message_filters::Subscriber<sensor_msgs::PointCloud2>> cloud_sub;
  std::unique_ptr<message_filters::Synchronizer<ApproxSyncPolicy>> sync;

  ros::Subscriber gps_sub;
  ros::Subscriber nmea_sub;
  ros::Subscriber navsat_sub;
  ros::Subscriber imu_sub;
  ros::Subscriber floor_sub;

  ros::Publisher markers_pub;

  std::string published_odom_topic;
  std::string map_frame_id;
  std::string odom_frame_id;

  std::mutex trans_odom2map_mutex;  // 一个互斥量,用来确保在多线程环境中对 trans_odom2map 变量的访问是线程安全的。
  Eigen::Matrix4f trans_odom2map;   // 表示从 里程计坐标系(odom frame)到地图坐标系(map frame)的变换矩阵       这个变换矩阵是跟机械臂变换矩阵类似的 齐次变换矩阵
  ros::Publisher odom2map_pub;      // 发布里程计到地图的变换信息

  std::string points_topic;  // 点云数据的ROS话题名称,SLAM系统将订阅这个话题来获取点云数据
  ros::Publisher read_until_pub;
  ros::Publisher map_points_pub;

  tf::TransformListener tf_listener;  // ROS的一个类,用来监听和查询不同坐标系之间的变换关系

  ros::ServiceServer load_service_server;      // 一个 ROS 服务服务器,用于加载系统数据或状态
  ros::ServiceServer dump_service_server;      // 一个 ROS 服务服务器,用于转储(dump)系统当前的状态或数据
  ros::ServiceServer save_map_service_server;  // 一个 ROS 服务服务器,用于保存地图

  // keyframe queue
  std::string base_frame_id;
  std::mutex keyframe_queue_mutex;           // std::mutex 是一种独占式互斥量
  std::deque<KeyFrame::Ptr> keyframe_queue;  // 关键帧队列

  // gps queue
  double gps_time_offset;
  double gps_edge_stddev_xy;
  double gps_edge_stddev_z;
  boost::optional<Eigen::Vector3d> zero_utm;
  std::mutex gps_queue_mutex;
  std::deque<geographic_msgs::GeoPointStampedConstPtr> gps_queue;

  // imu queue
  double imu_time_offset;                          // 用于对IMU数据时间戳进行校正或偏移
  bool enable_imu_orientation;                     // 控制是否启用IMU提供姿态信息
  double imu_orientation_edge_stddev;              // 用于指定IMU姿态数据的不确定性(标准差)
  bool enable_imu_acceleration;                    // 控制是否启用IMU提供的加速信息
  double imu_acceleration_edge_stddev;             // 指定IMU加速度信息的不确定性(标准差)
  std::mutex imu_queue_mutex;                      // 互斥量
  std::deque<sensor_msgs::ImuConstPtr> imu_queue;  // IMU队列

  // floor_coeffs queue
  double floor_edge_stddev;
  std::mutex floor_coeffs_queue_mutex;
  std::deque<hdl_graph_slam::FloorCoeffsConstPtr> floor_coeffs_queue;

  // for map cloud generation
  std::atomic_bool graph_updated;                          // 用于标识 位资图(pose graph)是否更新,确保多线程环境下的线程安全
  double map_cloud_resolution;                             // 控制生成的地图点云的分辨率
  std::mutex keyframes_snapshot_mutex;                     // 互斥量
  std::vector<KeyFrameSnapshot::Ptr> keyframes_snapshot;   // 用来存储关键帧的快照。这个变量很重要,发布 map_point 和 保存 pcd 地图都是用这个变量!!
  std::unique_ptr<MapCloudGenerator> map_cloud_generator;  // 生成点云地图的对象指针

  // graph slam
  // all the below members must be accessed after locking main_thread_mutex
  std::mutex main_thread_mutex;  // 互斥量

  int max_keyframes_per_update;             // 限制每次更新中可以处理的最大关键帧数量
  std::deque<KeyFrame::Ptr> new_keyframes;  // 存储新生成的关键帧指针

  g2o::VertexSE3* anchor_node;           // 一个指向 g2o::VertexSE3 类型的指针,表示位姿图中的一个特定节点(锚点节点)。锚点节点通常是用来约束图的某个部分,使得图优化时可以保持稳定。
                                         // 锚点的引入可以帮助优化算法更好地收敛到一个合理的解,特别是在 SLAM 系统中,锚点可以用于固定某个已知的位置(如起始位置或特征点)。
  g2o::EdgeSE3* anchor_edge;             // 一个指向 g2o::EdgeSE3 类型的指针,表示连接锚点节点和其他节点的边。这条边用于描述锚点与其他位姿节点之间的相对约束关系。
                                         // 在图优化中,这种约束有助于保持图的稳定性,并确保锚点的影响在优化过程中得以体现。
  g2o::VertexPlane* floor_plane_node;    // 用于表示地面平面节点的指针。
  std::vector<KeyFrame::Ptr> keyframes;  // 存储当前系统中的所有关键帧。
  std::unordered_map<ros::Time, KeyFrame::Ptr, RosTimeHash> keyframe_hash;  // 用于快速查找关键帧的哈希表
                                                                            // keyframe_hash 是一个无序映射,使用 ros::Time 作为键,以关键帧指针为值。此哈希表用于存储每个关键帧的时间戳与其对应的指针,便于快速查找特定时间的关键帧。
                                                                            // 这种数据结构可以加速对关键帧的查找过程,提高 SLAM 系统在处理数据时的效率,尤其是在需要频繁检索关键帧时。

  // std::unique_ptr:一种独占所有权的智能指针,意味着同一时间内只能有一个std::unique_ptr指向某个对象
  // 智能指针是一种自动管理动态分配内存的类,它们在对象不再使用时自动释放内存,从而帮助防止内存泄漏
  std::unique_ptr<GraphSLAM> graph_slam;  // 声明一个指向GraphSLAM 类型的智能指针 graph_slam
  std::unique_ptr<LoopDetector> loop_detector;
  std::unique_ptr<KeyframeUpdater> keyframe_updater;
  std::unique_ptr<NmeaSentenceParser> nmea_parser;

  std::unique_ptr<InformationMatrixCalculator> inf_calclator;
};

}  // namespace hdl_graph_slam

PLUGINLIB_EXPORT_CLASS(hdl_graph_slam::HdlGraphSlamNodelet, nodelet::Nodelet)