landmark是路标点，apriltag、二维码等强特征，static landmark 的作用是用于快速定位的。

对landmark的观测是相对于Cartographer的tracking frame，用户负责发布话题landmark，类型cartographer_ros_msgs/LandmarkList. cartographer订阅landmark之后，会发布话题/landmark_poses_list，类型visualization_msgs/MarkerArray，在rviz上显示出来，用于确认cartographer计算出landmark在map坐标系中的坐标。 Cartographer负责建立landmark的约束，与栅格地图一起优化。

机器人应当一直运动以产生trajectory优化，如果机器人不动，无法触发优化，landmark无法加入图优化框架。landmark的观测应当是相对于 tracking frame

Try the following to see an impressive effect (only for testing):
假如进行以下步骤:

1. 将 landmark 的权重设置的很大，比如 1e8
2. 观测一个id为0的 landmark
3. 把机器人开到另一个地方
4. 又一个id为0的 landmark

第二次观测显然是错误的，但是由于landmark的极高权重, 机器人位姿会跳跃

lua 设置

设置 use_landmarks = true，话题landmark应具备一定频率，默认要求最低10Hz。话题的消息可以是空的，但必须提供，因为Cartographer严格给传感器数据按时间排序，使landmarks deterministic. 但是也可以设置TRAJECTORY_BUILDER.collate_landmarks = false，这样landmark就是non-deterministic而且是非阻塞的。GPS也是同理。

landmarks_sampling_ratio = 1 landmark观测的权重比，当某个观测不太靠谱的时候，可以把1改为较小的权重，例如0.5，从而减小对整体优化的错误影响。

如果只获得landmark的相对位置，没有朝向，那就把landmark rotation weight设置为0

每个landmark的ID和坐标应当是unique而且是一一对应

源码

要知道tag在相机坐标系的位置，相机坐标系和base_link之间的关系

void Node::HandleLandmarkMessage(
    const int trajectory_id,  const std::string& sensor_id,
    const cartographer_ros_msgs::LandmarkList::ConstPtr&  msg)
{
  absl::MutexLock lock(&mutex_);
  if (!sensor_samplers_.at(trajectory_id).landmark_sampler.Pulse())
      return;

  map_builder_bridge_.sensor_bridge(trajectory_id)->HandleLandmarkMessage(sensor_id, msg);
  // 添加下面代码
  geometry_msgs::TransformStamped stamped_transform;

  stamped_transform.header.stamp = ros::Time::now();
  stamped_transform.header.frame_id = "base_link";
  stamped_transform.child_frame_id = "QR_Code";
  stamped_transform.transform =  ToGeometryMsgTransform(ToRigid3d(msg->landmarks[0].tracking_from_landmark_transform).inverse() );

  tf_broadcaster_.sendTransform(stamped_transform);
}

void Node::PublishLandmarkPosesList(
    const ::ros::WallTimerEvent& unused_timer_event)
{
   // landmark话题必须有订阅者，一般是rviz
  if (landmark_poses_list_publisher_.getNumSubscribers() > 0)
  {
    absl::MutexLock lock(&mutex_);
    landmark_poses_list_publisher_.publish(
        map_builder_bridge_.GetLandmarkPosesList() );
  }
}

visualization_msgs::MarkerArray MapBuilderBridge::GetLandmarkPosesList()
{
  visualization_msgs::MarkerArray landmark_poses_list;
  const std::map<std::string, Rigid3d> landmark_poses =
      map_builder_->pose_graph()->GetLandmarkPoses();
  LOG(INFO)<<" landmark pose size: "<< landmark_poses.size();

  for (const auto& id_to_pose : landmark_poses) {
    landmark_poses_list.markers.push_back(CreateLandmarkMarker(
        GetLandmarkIndex(id_to_pose.first, &landmark_to_index_),
        id_to_pose.second, node_options_.map_frame));
  }
  return landmark_poses_list;
}

publish_fake_random_landmarks.py

Cartographer提供这个文件发布假的landmark，不用仔细研究内容。 节点landmark_sampler，发布话题landmark，类型cartographer_ros_msgs/LandmarkList。

这个节点不是一直发布landmark，有的消息是空的

header: 
  seq: 546
  stamp: 
    secs: 1650511353
    nsecs:  89869022
  frame_id: ''
landmarks: []

有内容的消息是

header: 
  seq: 129
  stamp: 
    secs: 1650511071
    nsecs: 393965959
  frame_id: ''
landmarks: 
  - 
    id: "112b2"
    tracking_from_landmark_transform: 
      position: 
        x: 0.329674929632
        y: 0.601299422592
        z: 0.0813299308386
      orientation: 
        x: -0.881422944417
        y: 0.388576528204
        z: -0.260058416952
        w: 0.0668692347593
    translation_weight: 100000.0
    rotation_weight: 100000.0
  - 
    id: "32c2b"
    tracking_from_landmark_transform: 
      position: 
        x: 0.747269396167
        y: 0.256305093347
        z: 0.389155962116
      orientation: 
        x: -0.349830783746
        y: -0.891493805837
        z: -0.163724271997
        w: -0.236752148157
    translation_weight: 100000.0
    rotation_weight: 100000.0
  - 
    id: "2c33a"
    tracking_from_landmark_transform: 
      position: 
        x: 0.698094719569
        y: 0.886938424567
        z: 0.408522616781
      orientation: 
        x: -0.242061311724
        y: 0.292763269814
        z: 0.419295012755
        w: 0.824553019213
    translation_weight: 100000.0
    rotation_weight: 100000.0
  - 
    id: "c3b1c"
    tracking_from_landmark_transform: 
      position: 
        x: 0.703663491421
        y: 0.146106286628
        z: 0.589856404568
      orientation: 
        x: -0.0750314150542
        y: -0.80116295871
        z: 0.151662985651
        w: 0.574026601412
    translation_weight: 100000.0
    rotation_weight: 100000.0
  - 
    id: "c12aa"
    tracking_from_landmark_transform: 
      position: 
        x: 0.409531720117
        y: 0.286439128213
        z: 0.686302598836
      orientation: 
        x: -0.326127457574
        y: -0.221216268782
        z: -0.89427654661
        w: -0.212070040376
    translation_weight: 100000.0
    rotation_weight: 100000.0

参考:
Understanding the Landmark Observation Concept
Cartographer源码阅读——带Landmark的demo运行和测试