elirtf/opus
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

dev #17

Merged
elirtf merged 10 commits from dev into master 2026-07-07 13:56:57 +00:00
Conversation 0 Commits 10 Files changed 35 +1917 -510
elirtf commented 2026-07-07 13:56:51 +00:00
Owner
Copy link
No description provided. 
Caddyfile mirroring nginx/nginx.conf (security headers, forward_auth gate on
go2rtc, /go2rtc prefix strip + HLS cache rules, ONVIF discovery long timeouts,
static + SPA fallback). Added as a 'caddy' compose profile on host :8081 so it
runs side-by-side with nginx (:80) for A/B testing — nginx stays default and
untouched. Caddy gives automatic HTTPS and transparent websockets at ~5x less
config. Caddyfile passes 'caddy validate'. See docs/caddy-trial.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reusable single-frame latency benchmark + results. On real hardware: OpenVINO
gives YOLOv8n a 3.4x speedup over OpenCV-DNN (66ms->19.5ms), landing within ~7ms
of MobileNet-SSD@300 while detecting all people SSD missed. Verdict: the YOLO+
OpenVINO migration is worth it for the motion-gated single-inference use case.
Results: docs/research/detector-benchmark-2026-06-30.md

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The digest topics file + dated digests are now tracked so they stop getting
swept into unrelated commits. Added FFmpeg to the standing digest scope.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
v1: the discovery probe asks each camera's ONVIF Analytics service whether it
runs its own motion/object detection (classify_native_detection — pure, unit-
tested) and stores it (has_native_detection True|False|NULL + info). Surfaced as
a per-camera Detection badge (Native/Opus/Unknown) in Device Setup, with an
operator override via the camera update API. Deferred per spec: consuming
camera-side events, auto-skipping Opus detection for native cameras.

Spec: docs/superpowers/specs/2026-06-30-onvif-native-detection-design.md

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Rewrites ObjectDetector to run a YOLO ONNX through the OpenVINO runtime (CPU,
LATENCY hint) instead of MobileNet-SSD on OpenCV-DNN -- ~3.4x faster than YOLO
on OpenCV-DNN and far more accurate (benchmark:
docs/research/detector-benchmark-2026-06-30.md). COCO's 80 classes replace
VOC's 20. detect() does letterbox preprocess + YOLO decode + NMS
(cv2.dnn.NMSBoxes) + scale-back in pure, unit-tested helpers; keeps the
singleton/semaphore/fail-open contract and the detect() signature, so callers
are unchanged.

Model provisioning mirrors the old pattern: drop yolov8n.onnx in
OPUS_MODELS_DIR (yolo export model=yolov8n.pt format=onnx imgsz=640) or set
OPUS_YOLO_MODEL_URL; absent model / missing openvino -> motion-only. imgsz is
read from the model (env override for weak hosts). Adds openvino + numpy deps.

Note: COCO class names differ from VOC (motorbike->motorcycle, +truck);
existing object_detect_classes settings using VOC-only names may need updating.
Validated end-to-end on the real model (3 persons + 1 bus on a test image).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Detector dropdown exposed backend internals (opencv/KNN/MOG2/Stub); replace
it with a simple Motion detection On/Off (On=KNN default, Off=Stub). Removed the
Advanced section (stagger delay + raw per-detector tuning params) — too granular
and never touched; those settings keep their defaults, just no longer shown.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
13 runbook skills distilled from the repo: architecture contract, change
control, build/env, operations, validation, debugging, failure archaeology,
config catalog, detection, streaming, recording/storage, migrations, docs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
elirtf merged commit e0edb7865a into master 2026-07-07 13:56:57 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
elirtf/opus!17
No description provided.