Give box/footprint/polygon filters unique private node names - #266
Give box/footprint/polygon filters unique private node names#266rafal-gorecki wants to merge 2 commits into
Conversation
These filters spin up their own internal node for TF and (for polygon) pub/sub/param-callback needs. That internal node used a fixed name, so a process-wide "-r __node:=<name>" remap (e.g. Node(name="laser_filter") in a launch file) silently renamed it too, colliding with the parent filter-chain node under the same name and breaking rosout logging. Give each internal node a name unique to the filter instance, pinned via its own local "-r __node:=<name>" argument so it survives a process-wide bare remap. Fixes ros-perception#214
|
Tick the box to add this pull request to the merge queue (same as
|
box_filter, footprint_filter and polygon_filter each construct their TransformListener with the buffer-only convenience constructor, which spins up yet another hidden node (and dedicated thread) purely to subscribe to /tf and /tf_static, on top of the private node these filters already carry. Pass that private node through instead, so there's one node per filter, not two, matching the same cleanup already done for the top-level filter-chain node in 1801589 ("Re-use the main filter node for the TransformListener").
|
A note on why this also drops These classes never actually engage the lifecycle state machine — no Despite that, the resulting node still advertises itself as a fully manageable lifecycle node ( |
|
@jonbinney please review |
1 participant
Fixes #214.
box_filter,footprint_filterandpolygon_filtereach spin up their owninternal node (for TF, and for polygon_filter also pub/sub/param callbacks).
That internal node used a fixed name, so a process-wide bare
-r __node:=<name>remap (e.g.Node(name="laser_filter")in a launch file)silently renames it too, colliding with the parent filter-chain node and
producing
Publisher already registered for node name: 'laser_filter',exactly as reported here and in #214 (comment).
Each internal node now gets a name unique to the filter instance, pinned via
its own local
-r __node:=<name>argument — that survives a process-widebare remap since node-specific rules take precedence over the wildcard one.
This follows the same approach @jonbinney used in #248 for box_filter/
polygon_filter (opened against #231), and extends it to footprint_filter,
which #248 doesn't cover but is affected by the same root cause.
Tested against the exact repro from the issue (scan_to_scan_filter_chain,
--remap __node:=laser_filter, box_filter via params file): the"Publisher already registered" warning and the node-name collision are
gone, and the box filter's internal node keeps a unique name in
ros2 node list.