The object_merger package fuses detected objects from multiple sources using data association. It relies on the same SSP (Successive Shortest Path) algorithm used in multi_object_tracker for solving the minimum-cost flow data association problem. The cost between objects is computed from their distance, and gates (maximum distance, maximum area, minimum area) reset the cost when violated.

Node Implementation

Enum: MSG_COV_IDX

Defines indices for all covariance entries between state variables.

Class: ObjectFusionNode

The main node class handling merging logic.

Function: IsOverlapWithUnknown

Checks if an Unknown object overlaps with a known object. It takes a known object, computes distance, and then evaluates overlap based on distance, precision, recall, and Generalized IoU thresholds.

Function: BuildClassThresholdMap

Converts a list of distance thresholds (from parameter distance_threshold_list) in to a map associating each class label with its threshold. This map is consumed by IsOverlapWithUnknown.

Constructor: ObjectFusionNode::ObjectFusionNode

The node subscribes to two topics, each carrying detection results to be merged:

Topic	Type	Description
input/object0	autoware_auto_perception_msgs::msg::DetectedObjects	First set of detection objects
input/object1	autoware_auto_perception_msgs::msg::DetectedObjects	Second set of detection objects

Parameters loaded:

Name	Type	Description
can_assign_matrix	double	Assignment eligibility table
max_dist_matrix	double	Maximum distance gate
max_area_matrix	double	Maximum area gate
min_area_matrix	double	Minimum area gate
max_rad_matrix	double	Maximum angle gate
base_link_frame_id	string	Frame for association
distance_threshold_list	std::vector<double>	Per-class overlap distance thresholds
generalized_iou_threshold	double	GIoU threshold for overlap

The constructor sets up a synchronized callback using sync_:

sync_(SyncPolicy(10), sub0_, sub1_);
sync_.registerCallback(std::bind(&ObjectFusionNode::OnObjectsReceived, this, _1, _2));

It also instantiates a DataAssociation object and builds the overlap threshold map:

overlap_params_.distance_threshold_map =
  BuildClassThresholdMap(declare_parameter<std::vector<double>>("distance_threshold_list"));

Callback: ObjectFusionNode::OnObjectsReceived

Both object lists are first transformed to the base_link frame. Then data association is performed:

std::unordered_map<int, int> direct_assignment, reverse_assignment;
const auto & objects0 = transformed_objects0.objects;
const auto & objects1 = transformed_objects1.objects;
Eigen::MatrixXd score_matrix =
  data_assoc_->ComputeScoreMatrix(transformed_objects1, transformed_objects0);
data_assoc_->Assign(score_matrix, direct_assignment, reverse_assignment);

For each matched pair, the output object is selected based on the configured PriorityMode:

for (size_t i = 0; i < objects0.size(); ++i) {
  const auto & obj0 = objects0[i];
  if (direct_assignment.contains(i)) {
    const auto & obj1 = objects1[direct_assignment[i]];
    switch (priority_) {
      case PriorityMode::Source0:
        output.objects.push_back(obj0);
        break;
      case PriorityMode::Source1:
        output.objects.push_back(obj1);
        break;
      case PriorityMode::Confidence:
        if (obj1.existence_probability <= obj0.existence_probability)
          output.objects.push_back(obj0);
        else
          output.objects.push_back(obj1);
        break;
    }
  } else {
    output.objects.push_back(obj0);
  }
}

Unmatched objects from the second list are appended:

for (size_t j = 0; j < objects1.size(); ++j) {
  if (!reverse_assignment.contains(j))
    output.objects.push_back(objects1[j]);
}

Finally, overlapping Unknown objects are filtered out by scanning the merged list with IsOverlapWithUnknown and removing those that overlap with known objects.