recamera-lite

Two-tier, salience-gated vision for the Seeed reCamera.

A cheap, always-on on-device NPU detector watches the scene. Only when something salient appears does the camera (a) fire a small UDP event and (b) hand off the triggering frame to a local GPU running a heavier YOLO that confirms and upscales the detection. When nothing is happening, the edge is silent and the GPU does zero work.

   reCamera (edge)                              your server (GPU)
 ┌────────────────────┐                       ┌─────────────────────────┐
 │ camera → NPU yolo  │   UDP event :23906    │ event listener          │
 │ (cheap, always on) │ ────────────────────▶ │   ▼ while "active"       │
 │ salience + threat  │                       │ pull gated frame (HTTP) │
 │ writes gated JPEG  │ ◀──── GET /vis/x.jpg ──│   ▼                     │
 │ to tmpfs (lighttpd)│                       │ GPU YOLO confirm        │
 └────────────────────┘                       │   ▼ dashboard + API     │
                                              └─────────────────────────┘
   FPS = threat level: capture rate scales with how interesting the scene is;
   idle ⇒ ~0 fps ⇒ no traffic, no GPU inference.

This repo has two halves:

Dir	What it is
edge/	the lite OS daemon for the reCamera — the native vision daemon (RTSP + NPU salience + threat-gated JPEG + UDP events), an installer, a config file, and the on-device model-build recipe
server/	the GPU server — one self-contained Python service: UDP event listener + gated-frame pull + in-process Ultralytics YOLO on your GPU + a live dashboard and JSON API

See docs/ARCHITECTURE.md for the design and the full wire protocol.

Quick start — the server (no camera needed to try it)

cd server
python3 -m venv .venv && . .venv/bin/activate
# install a CUDA torch for your driver first (see requirements.txt), then:
pip install -r requirements.txt
python3 server.py                      # http://localhost:8092/

Test the GPU tier directly with any JPEG — no camera required:

curl --data-binary @photo.jpg -H 'Content-Type: image/jpeg' http://localhost:8092/detect

Or simulate a whole edge device (events + a gated frame) with no hardware:

sudo python3 tools/simulate_edge.py --img photo.jpg --server 127.0.0.1   # then watch the dashboard

Quick start — the edge

The reCamera daemon is C++ and cross-compiled against the Seeed SG2002 SDK (not vendored here — multi-GB). Build it, then deploy:

cd edge/src && make                    # needs SG200X_SDK_PATH + the musl toolchain (see src/Makefile)
cd .. && ./install.sh <device-ip> src/vision model-build/yolo11n_ours.cvimodel
# then set SERVER=<your-gpu-box>:23906 in /userdata/recamera-lite.conf on the device

Full edge notes (sensors, model build, config, gotchas): edge/README.md.

Configuration at a glance

• Server (env vars): YOLO_MODEL (n/s/m/l/x), YOLO_IMGSZ, YOLO_CONF, EVT_PORT, HTTP_PORT, PULL_CAP, CONFIRM_CLASSES, SAVE_DIR. See server/README.md.
• Edge (/userdata/recamera-lite.conf): SERVER, CAM_NAME, MODEL, SALCONF, FPSMAX. See edge/README.md.

License

Apache-2.0. The reCamera edge daemon links the Seeed/Sophgo CVITEK MPI + TPU runtime SDKs, which carry their own (vendor) licenses — install those from Seeed; they are not redistributed here.

Credits

Distilled from a private home-vision build into a clean, general-purpose release. Targets the reCamera Gen1 (SG2002/CV181x, OV5647) and reCamera HQ (GC2053). YOLO models via Ultralytics; on-device INT8 compile via TPU-MLIR.