github BennoBaer-dev/suslik v0.1.1.110
suslik 0.1.1.110

57 minutes ago

Test release. 0.1.1.104 to 0.1.1.109 were never published; their changes are part of this version.

  • Person recognition (preview) is rebuilt and optional. It learns body images per person from videos in which the face path recognized exactly that person. An event without a recognized face and with exactly one person then counts as a vote when its body picture is close enough to one of them; by default a name counts when one pass gives the same person three votes (koerper_stimmen). Details: docs/person-recognition.md and the chapter Person recognition in Read me first.
  • Only videos analyzed with this version can give body images; older videos don't count. The path does nothing until there is at least one valid body image.
  • New page Person → Body images: start the learning run, choose how many videos per person to check, accept or discard new images (by default only images you accept count) and see the stock per person. koerper_lauf sets when the run starts: hand (by the button, the default), taeglich (once a day with the maintenance at 06:00) or sofort (right after each suitable video).
  • A body name is shown on Today, Passes and Appearances. Five switches, all off by default, let it go further: koerper_meldet_frigate (sub label, needs sub_label_schreiben), koerper_meldet_pushover, koerper_meldet_telegram, koerper_meldet_ha (MQTT topic koerper_erkannt) and koerper_meldet_anwesenheit (marks on the Presence page and in the hour bar on Today, nothing sent outside). Pushover, Telegram and Home Assistant send at most once per pass and person, and stay silent when the face path confirmed the same person within anwesenheit_cooldown.
  • Recognition page: the body card shows these switches and the choice Start learning run directly. The save button is now a bar at the bottom of the page, grey with "All saved" until you change something, then highlighted with "Not saved, saving restarts the service".
  • New settings, all on the Advanced page: koerper_auswahl_stimmen (6), koerper_bilder_je_video (3), koerper_anker_min (0.5), koerper_sichtpruefung (on), koerper_lauf (hand), koerper_max_tage (10), koerper_max_je_person (50), koerper_schwelle (0.65), koerper_marge (0.02), koerper_stimmen (3) and the five switches above (off). None of the defaults is measured on more than one system; see docs/configuration.md.
  • The old person-learn run is removed: its pages, its model and training pictures, the alerts with the note "person recognition, not face" and the MQTT message person_erkennung. The learned faces stay. If you used it, person recognition has to learn again. The maintenance step "data cleanup" moves the folder personlern/ into the trash, which is emptied 7 days later.
  • Vision detect is parked: it is switched off and cannot be switched on in this version, a saved "on" is set to off once at the first start, and the Vision pages say why.
  • Telegram: with telegram_modus set to aus (off, the default), nothing is sent over Telegram any more, including fault messages and the messages and videos of the live watchers. The channel test on the page still works.
  • Body images live in refquelle/koerperbilder/ in the data folder and are part of the full backup and of /support/faces; the support area /support/personlern is gone. The images no longer carry the DINOv2 model file.

gpu-legacy: not load-tested before this release (no test machine for this variant).
rocm: not load-tested before this release (no test machine for this variant); a field installation last confirmed this variant on 0.1.0.511.

Known issues, accepted for this release:

  • cpu: in the first load test the live engine's warning counter in /health was two lines behind the log at the moment of the last snapshot (40 against 42), because the live engine wrote two more warnings after it; every event was booked once with the persons of the reference run.

Don't miss a new suslik release

NewReleases is sending notifications on new releases.