radar_object_clustering

This package contains a radar object clustering for autoware_perception_msgs/msg/DetectedObject input.

This package can make clustered objects from radar DetectedObjects, the objects which is converted from RadarTracks by radar_tracks_msgs_converter and is processed by noise filter. In other word, this package can combine multiple radar detections from one object into one and adjust class and size.

Design

Background

In radars with object output, there are cases that multiple detection results are obtained from one object, especially for large vehicles such as trucks and trailers. Its multiple detection results cause separation of objects in tracking module. Therefore, by this package the multiple detection results are clustered into one object in advance.

Algorithm

• 1. Sort by distance from base_link

At first, to prevent changing the result from depending on the order of objects in DetectedObjects, input objects are sorted by distance from base_link. In addition, to apply matching in closeness order considering occlusion, objects are sorted in order of short distance in advance.

• 1. Clustering

If two radar objects are near, and yaw angle direction and velocity between two radar objects is similar (the degree of these is defined by parameters), then these are clustered. Note that radar characteristic affect parameters for this matching. For example, if resolution of range distance or angle is low and accuracy of velocity is high, then distance_threshold parameter should be bigger and should set matching that strongly looks at velocity similarity.

After grouping for all radar objects, if multiple radar objects are grouping, the kinematics of the new clustered object is calculated from average of that and label and shape of the new clustered object is calculated from top confidence in radar objects.

• 1. Fixed label correction

When the label information from radar outputs lack accuracy, is_fixed_label parameter is recommended to set true. If the parameter is true, the label of a clustered object is overwritten by the label set by fixed_label parameter. If this package use for faraway dynamic object detection with radar, the parameter is recommended to set to VEHICLE.

• 1. Fixed size correction

When the size information from radar outputs lack accuracy, is_fixed_size parameter is recommended to set true. If the parameter is true, the size of a clustered object is overwritten by the label set by size_x, size_y, and size_z parameters. If this package use for faraway dynamic object detection with radar, the parameter is recommended to set to size_x, size_y, size_z, as average of vehicle size. Note that to use for multi_objects_tracker, the size parameters need to exceed min_area_matrix parameters of it.

Limitation

For now, size estimation for clustered object is not implemented. So is_fixed_size parameter is recommended to set true, and size parameters is recommended to set to value near to average size of vehicles.

Interface

Input

• ~/input/objects (autoware_perception_msgs/msg/DetectedObjects.msg)
  • Radar objects

Output

• ~/output/objects (autoware_perception_msgs/msg/DetectedObjects.msg)
  • Output objects

Parameter

• angle_threshold (double) [rad]
  • Default parameter is 0.174.
• distance_threshold (double) [m]
  • Default parameter is 4.0.
• velocity_threshold (double) [m/s]
  • Default parameter is 2.0.

These parameter are thresholds for angle, distance, and velocity to judge whether radar detections come from one object in "clustering" processing, which is written in detail at algorithm section. If all of the difference in angle/distance/velocity from two objects is less than the thresholds, then the two objects are merged to one clustered object. If these parameter is larger, more objects are merged to one clustered object.

These are used in isSameObject function as below.

bool RadarObjectClusteringNode::isSameObject(
  const DetectedObject & object_1, const DetectedObject & object_2)
{
  const double angle_diff = std::abs(autoware::universe_utils::normalizeRadian(
    tf2::getYaw(object_1.kinematics.pose_with_covariance.pose.orientation) -
    tf2::getYaw(object_2.kinematics.pose_with_covariance.pose.orientation)));
  const double velocity_diff = std::abs(
    object_1.kinematics.twist_with_covariance.twist.linear.x -
    object_2.kinematics.twist_with_covariance.twist.linear.x);
  const double distance = autoware::universe_utils::calcDistance2d(
    object_1.kinematics.pose_with_covariance.pose.position,
    object_2.kinematics.pose_with_covariance.pose.position);

  if (
    distance < node_param_.distance_threshold && angle_diff < node_param_.angle_threshold &&
    velocity_diff < node_param_.velocity_threshold) {
    return true;
  } else {
    return false;
  }
}

• is_fixed_label (bool)
  • Default parameter is false.
• fixed_label (string)
  • Default parameter is "UNKNOWN".

is_fixed_label is the flag to use fixed label. If it is true, the label of a clustered object is overwritten by the label set by fixed_label parameter. If the radar objects do not have label information, then it is recommended to use fixed label.

• is_fixed_size (bool)
  • Default parameter is false.
• size_x (double) [m]
  • Default parameter is 4.0.
• size_y (double) [m]
  • Default parameter is 1.5.
• size_z (double) [m]
  • Default parameter is 1.5.

is_fixed_size is the flag to use fixed size parameters. If it is true, the size of a clustered object is overwritten by the label set by size_x, size_y, and size_z parameters.