Filters

Notes

The repository package is named bb_filters.

Clustering APIs

Two APIs are available:

• ClusterTfAction and ClusterTfSrv sample transforms from TF and broadcast a stable output transform.
• ClusterPosesAction and ClusterPosesSrv cluster pose topics directly and return ClusterPoseResultArray.

Both use spatial clustering to reject outliers before selecting a stable result.

Operational notes

• Use simulated time when clustering simulated messages or transforms.
• A TF clustering request needs a live broadcaster for every input child frame. Repeated LookupException errors normally mean the upstream pose estimator is not publishing.

Cluster Poses vs Transforms

We implement separate cluster poses nodes to:

• Control time alignment explicitly. We can ensure each detected pose is paired with an odometry sample within a chosen tolerance, rather than relying on latest available transforms.
• Produce physically meaningful clustering inputs. Each data point corresponds to a real detection event transformed using the odometry state at approximately the same time, rather than "virtual" points created by mixing a detection at time t1 with odometry at time t2. (When clustering transforms, you may feed in multiple copies of the same detection transformed with different odometry readings.)
• Achieve higher efficiency from not having to subscribe to a (potentially) high frequency dynamic transforms topic /tf and not having to use tf lookups. (But this is balanced out by the overhead from managing the time synchronization ourselves instead of using a tf2_ros.Buffer.)

Cluster Poses implementation considerations

• TOCTOU race condition see here
• Performance drawback with MultithreadedExecutors see here