沉默杀手HomeArchivesTagsCategories
landmark的前端处理
|激光SLAMCartographer源码解读|
Word count: 653|Reading time: 3 min

landmark就是添加了一个观测，让定位更鲁棒。

In order to make Cartographer calculations deterministic we strictly time order sensor data using the collator. But you only know in what order to publish data when you have already seen data from all other sources. Sparse sources pose a problem to that strategy.

如果观测是空的，应当发布空的 landmark messages to unblock the queue. 对于fixed frame poses (GPS)也适用。

前端的处理

node.cc 层

1
2
3
4
// landmark_poses_list 话题
  landmark_poses_list_publisher_ =
      node_handle_.advertise<::visualization_msgs::MarkerArray>(
          kLandmarkPosesListTopic, kLatestOnlyPublisherQueueSize);

发布话题：

1
2
3
4
5
6
7
8
void Node::PublishLandmarkPosesList(const ::ros::WallTimerEvent& unused_timer_event)
{
  if (landmark_poses_list_publisher_.getNumSubscribers() > 0)
  {
      absl::MutexLock lock(&mutex_);
      landmark_poses_list_publisher_.publish(map_builder_bridge_.GetLandmarkPosesList() );
  }
}

处理传感器数据

1
2
3
4
5
6
7
8
9
10
11
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;
  CHECK(msg->landmarks.size() != 0) << "No Landmarks !";

  map_builder_bridge_.sensor_bridge(trajectory_id)->HandleLandmarkMessage(sensor_id, msg);
}

sensor_bridge.cc 层

用到的类型

1
2
3
4
5
6
7
8
9
10
11
12
struct LandmarkData {
  common::Time time;
  std::vector<LandmarkObservation> landmark_observations;
};

struct LandmarkObservation
{
  std::string id;
  transform::Rigid3d landmark_to_tracking_transform;
  double translation_weight;
  double rotation_weight;
};

landmark的观测是相对于 tracking frame

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
void SensorBridge::HandleLandmarkMessage(
    const std::string& sensor_id,
    const cartographer_ros_msgs::LandmarkList::ConstPtr& msg)
{
  auto landmark_data = ToLandmarkData(*msg);
  auto tracking_from_landmark_sensor = tf_bridge_.LookupToTracking(
      landmark_data.time, CheckNoLeadingSlash(msg->header.frame_id));
  // 把相机对landmark的观测，转换到 tracking 坐标系
  if (tracking_from_landmark_sensor != nullptr)
  {
      for (auto& observation : landmark_data.landmark_observations)
      {
        observation.landmark_to_tracking_transform =
            *tracking_from_landmark_sensor *
            observation.landmark_to_tracking_transform;
      }
  }
  // 进入传感器数据融合阶段
  trajectory_builder_->AddSensorData(sensor_id, landmark_data);
}

简单的类型转换而已

1
2
3
4
5
6
7
8
9
10
11
12
13
14
LandmarkData ToLandmarkData(const LandmarkList& landmark_list)
{
    LandmarkData landmark_data;
    landmark_data.time = FromRos(landmark_list.header.stamp);
    for (const LandmarkEntry& entry : landmark_list.landmark)
    {
        landmark_data.landmark_observations.push_back(
        {entry.id, ToRigid3d(entry.tracking_from_landmark_transform),
         entry.translation_weight, 10, entry.type});
         //LOG(INFO) << "find landmark x: " << entry.tracking_from_landmark_transform.position.x;
         //LOG(INFO) << "find landmark y: " << entry.tracking_from_landmark_transform.position.y;
    }
    return landmark_data;
}

数据融合阶段

CollatedTrajectoryBuilder类

1
2
3
4
5
6
7
8
9
10
11
12
void AddSensorData(const std::string& sensor_id,
                   const sensor::LandmarkData& landmark_data) override
{
    if (collate_landmarks_)
    {
      AddData(sensor::MakeDispatchable(sensor_id, landmark_data));
      return;
    }
    // wrapped_trajectory_builder_ 赋值在 CollatedTrajectoryBuilder 构造函数
    // 应当为map_builder_->GetTrajectoryBuilder(trajectory_id)
    wrapped_trajectory_builder_->AddSensorData(sensor_id, landmark_data);
}

  • collate_landmarks的设置问题

在CollatedTrajectoryBuilder的构造函数里体现，比较坑的是设置为on或off，不是true, false

1
2
3
4
5
if (sensor_id.type == SensorId::SensorType::LANDMARK &&
    !collate_landmarks_)
{
    continue;
}

如果提供landmark的观测不低于10 Hz，那么可以设置TRAJECTORY_BUILDER.collate_landmarks = on. Cartographer将deterministically运行(对于给定的bag, 使用offline node每次获得的结果是一样的). 如果collate_landmarks = off， landmark observations将跳过sensor queue (so they do not block it if they are sparse) and are injected directly into the pose graph, which is not deterministic.

但是看到另一种说法，在 #1224更新版本，作者提出如果有landmark，collate_landmarks就要为 false。这个再观察

接下来进入GlobalTrajectoryBuilder

1
2
3
4
5
void AddSensorData(const std::string& sensor_id,
                   const sensor::LandmarkData& landmark_data) override
{
   pose_graph_->AddLandmarkData(trajectory_id_, landmark_data);
}

这里可以看到landmark数据直接进后端，前端没有它的事了