devcam 0.1.0

Local camera feeds for development with HTTP, RTSP, and optional YOLO human detection.

devcam

Easy local camera feeds for development and testing. devcam can expose a camera as HTTP MJPEG, publish RTSP, and optionally run YOLO human detection from the same web UI.

Components

Service	Port	Description
python -m devcam http	1080	Web UI, HTTP MJPEG stream, snapshots, camera controls, optional human detection
python -m devcam rtsp	1081	RTSP publisher (requires ffmpeg + mediamtx)
human_detector/detector.py	1082	Legacy standalone YOLOv8n person detection API

Quick Start

pip install -e .
python -m devcam http

Install YOLO human detection support with the optional vision extra:

pip install -e ".[vision]"

Open http://localhost:1080/ to view the stream, switch cameras, change resolution/FPS, mirror webcam view, turn capture on or off, and enable human detection. Detection is lazy-loaded: the app starts without loading YOLO, and the model loads only when detection is enabled or /api/detection/run is called.

The legacy entrypoint devcam CLI command also works:

devcam http

Camera Server

Captures webcam and broadcasts as HTTP MJPEG or RTSP. Pick one at launch.

python -m devcam {http,rtsp} [--port PORT] [--camera INDEX] [--resolution WxH] [--fps FPS] [--backend {auto,cv2,zed}]

• HTTP mode → http://localhost:1080/ (web viewer), /stream (MJPEG), /snapshot (single JPEG)
• RTSP mode → rtsp://localhost:1081/cam (requires ffmpeg + mediamtx)

The HTTP web viewer includes controls for camera selection, resolution, FPS, mirrored webcam display, turning capture on or off, enabling human detection, changing confidence threshold, changing detection rate up to 20 Hz, and drawing detection boxes over the live stream.

Flag	Default	Description
protocol	(required)	http or rtsp
--port	1080 / 1081	Stream port
--camera	0	Camera index
--resolution	auto	WxH e.g. 1280x720
--fps	30	Capture and stream frame rate
--backend	auto	auto (ZED first, then OpenCV), cv2 (V4L2/USB), zed (ZED SDK)

HTTP Control API

Endpoint	Method	Description
/api/status	GET	Current backend, selected camera, requested mode, actual mode, and capture state
/api/cameras	GET	Available cameras for the active backend
/api/modes	GET	Common resolution and FPS choices for the UI
/api/config	POST	Start, stop, mirror, or switch camera settings

Example:

curl -X POST http://localhost:1080/api/config \
  -H 'Content-Type: application/json' \
  -d '{"enabled":true,"camera":0,"width":1280,"height":720,"fps":15,"mirror":true}'

Human Detection API

Human detection runs in the same Flask process against the latest in-memory camera frame. It does not require running the old detector service in a second terminal.

Endpoint	Method	Description
/api/detection/status	GET	Detection availability, enabled state, confidence threshold, interval, model loaded state, and errors
/api/detection/config	POST	Enable/disable detection and update settings
/api/detection/run	GET/POST	Run YOLO person detection on the latest frame

Configuration example:

curl -X POST http://localhost:1080/api/detection/config \
  -H 'Content-Type: application/json' \
  -d '{"enabled":true,"confidence":0.55,"interval_ms":50}'

interval_ms controls detection cadence. The UI slider supports 1-20 Hz, where 20 Hz is 50 ms.

Detection response shape:

{
  "enabled": true,
  "available": true,
  "human_detected": true,
  "confidence": 0.87,
  "confidence_threshold": 0.55,
  "count": 1,
  "frame_width": 1920,
  "frame_height": 1080,
  "latency_ms": 42.1,
  "detections": [
    {"confidence": 0.87, "bbox": [420.0, 160.0, 780.0, 940.0]}
  ]
}

If ultralytics is not installed, camera streaming still works. Detection endpoints return available: false with a clear error.

Python Package

devcam is packaged with pyproject.toml. For local development:

pip install -e ".[vision]"
devcam http

Build distribution artifacts:

pip install -e ".[dev]"
python -m build

Before publishing, verify the package metadata and artifacts:

python -m twine check dist/*

Publish to TestPyPI first:

python -m twine upload --repository testpypi dist/*

Then publish to PyPI when ready:

python -m twine upload dist/*

Before a real PyPI release, confirm the devcam package name is available, choose a license intentionally, and consider adding a LICENSE file.

Camera Backends

• cv2 — OpenCV V4L2 for USB webcams
• zed — ZED SDK for ZED X / ZED X Mini on Jetson (requires pyzed)
• auto (default) — tries ZED first, falls back to OpenCV

Jetson / ZED X Notes

• If ZED cameras show as "NOT AVAILABLE", restart the daemon: sudo systemctl restart zed_x_daemon
• Resolution is always opened at the camera's native mode; --resolution downscales in software
• --fps 5 or --fps 15 recommended for low-bandwidth use cases

RTSP Setup

One-time setup on macOS/Linux:

# macOS: brew install ffmpeg
# Linux: sudo apt install ffmpeg
bash setup_mediamtx.sh   # downloads mediamtx for this OS/CPU

One-time setup on Windows PowerShell:

Windows setup has not been tested or validated yet.

winget install Gyan.FFmpeg
.\setup_mediamtx.ps1    # downloads mediamtx for Windows/CPU

Each time you use RTSP on macOS/Linux:

./mediamtx               # run in a separate terminal
python -m devcam rtsp

Each time you use RTSP on Windows PowerShell:

.\mediamtx.exe           # run in a separate terminal
python -m devcam rtsp

human_detector

REST API for person detection using YOLOv8n (~6MB model, CPU-only, auto-downloaded).

python human_detector/detector.py [--port PORT] [--preload]

Endpoint	Method	Input	Response
/detect?url=...	GET	Image URL	{human_detected, confidence, count, detections}
/detect	POST	File upload / raw bytes / JSON {"url":"..."}	Same
/health	GET	—	{"status": "ok"}

Benchmark

python human_detector/benchmark.py -n 50

Measures detection throughput (Hz) using a pre-fetched snapshot.

macOS Notes

• Camera permission: System Settings → Privacy & Security → Camera → Terminal
• All ports are unprivileged (≥1024) — no sudo needed