Changelog

Global Association Multicamera Tracking filter release notes

v0.1.14 - 2026-10-22Direct link to v0.1.14 - 2026-10-22

With per_class_global_ids, each class publishes its own global id sequence. The registry key stays unique, so two classes can both be 1. A class flicker on the same local track keeps the original number and class. The published id is the bare per-class number, so global_id_prefix still forms the registry key and is not part of global_id.
The drawn global id sits on the top-right corner of the box.

With per_class_global_ids on, a blank class is its own bucket and does not take a labeled track's number.
Distance checks use the bbox last seen on the current camera, so a cam_2 track does not lose its id when cam_1 writes a distant box.
When source_allow_same_camera_reassociation is true, a source camera can re-acquire a global id that an overlapping camera still holds. The new box must sit within same_camera_reassociation_max_center_distance of the box last stored for the source camera, the gap since that overlapping camera last saw the entry must fall inside [same_camera_reassociation_after_seconds, same_camera_reassociation_max_seconds], and the association score must reach match_threshold. A gap inside the window whose score is still below the threshold mints a new id. A distant track on the same camera still mints. A local-id switch on the source camera also still mints when that id was last seen on the source camera itself: the flag does not turn off always_create_new_on_source_camera for those entries. With the flag off, the source camera keeps minting.

With per_class_global_ids off, a missing class no longer matches a labeled one.
Bump the openfilter dependency to 1.5.0
Bump the openfilter dependency to 1.5.1

Build the filter image on openfilter-base:py3.14 (was py3.11). The published wheel supports Python 3.14, so the image now ships 3.14. Running on 3.10–3.13 is unaffected.

Update the openfilter dependency to 1.3.0
Add Python 3.13 and 3.14 support: raise the requires-python ceiling to <3.15; the CI test matrix now runs 3.10–3.14.

Build the image on openfilter-base (weekly apt-upgraded python-slim) instead of a stale python:X.Y.Z-slim pin, clearing the OS-package CVEs the pin carried.
Update the openfilter dependency to 1.2.2

Update openfilter[all] to >=1.2.1
Replaced the exact opencv-python-headless==4.13.0.90 pin with the bounded range >=4.13,<6. The filter only uses rectangle, putText, getTextSize and FONT_HERSHEY_SIMPLEX, and opencv also arrives transitively through openfilter, so an exact pin here only creates a conflict the moment that transitive constraint moves.
Pin the Docker base to python:3.11.12-slim.
Refresh the docker-compose.yaml openfilter utility image tags to 1.2.1 and pin the filter's own image to the release version.

Fixed global_id to be sticky on single camera tracks, preventing id freezing when cam2 is added to the association

Add QUICKSTART.md with full two-camera pipeline examples, association tuning guide, and containment tracking section
Expand .env.example to cover all FILTER_* parameters including gating, weights, and containment tracking
Rewrite README to remove stale template text and document the full pipeline, output contract, and configuration