Apppublic

naveenkm13/occupancyos

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes

App README

CCTV Workplace Occupancy Detection

YOLOv8l + OpenCV + FastAPI + React real-time workplace / classroom / coworking occupancy analytics. Detects people and chairs with COCO-pretrained YOLOv8l and marks each chair as occupied/free based on person-on-chair overlap. Broadcasts to a live dashboard over WebSockets.

Two detection modes (switchable via DETECTOR_MODE):

—`chair` (default) — dynamic chair detection; no setup, works on any camera angle.
—`zone` — predefined desk rectangles in pixel space (config.DESK_ZONES).

VIDEO FILE / RTSP CCTV STREAM
        ↓ OpenCV
   YOLOv8l Person Detection
        ↓
   Desk Zone Mapping  (centroid → rectangle)
        ↓
   Occupancy Analytics
        ↓ FastAPI WebSocket
   React Dashboard

Features

—YOLOv8l person detection (COCO-pretrained or fine-tuned)
—Centroid-based desk-zone occupancy (no chair detection)
—Video file, webcam, and RTSP support
—FastAPI + WebSocket real-time broadcast (~10 Hz)
—MJPEG snapshot stream for live preview
—Analytics: utilization %, peak/avg, per-desk dwell, alerts, CSV export
—React + Vite + Tailwind dashboard with charts, dark glassmorphism UI
—Full training pipeline (extract → annotate → split → train → validate → export)

Project layout

backend/
  app/           FastAPI app + pipeline
  detection/    YOLOv8l + occupancy engine
  analytics/    history, alerts, CSV export
  api/          REST endpoints
  streams/      OpenCV video sources (file / webcam / RTSP)
  websocket/    connection manager
  training/     prepare_dataset, train, validate
  scripts/      video_inference, websocket_server, extract_frames
  models/       weights (yolov8l.pt, best.pt)
frontend/        React + Vite + Tailwind dashboard

Quick start (backend)

1. Python environment

bash

python -m venv .venv
.\.venv\Scripts\activate          # Windows
# source .venv/bin/activate       # Linux/Mac
pip install -r requirements.txt

2. CUDA-enabled PyTorch (recommended)

The default pip install of torch may install the CPU build. To use GPU:

bash

pip install --index-url https://download.pytorch.org/whl/cu121 \
    torch torchvision

Verify:

bash

python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"

3. Place a model

The first run will auto-download yolov8l.pt via Ultralytics. Or place your fine-tuned best.pt at backend/models/best.pt and set:

bash

set YOLO_WEIGHTS=backend\models\best.pt   # Windows
# export YOLO_WEIGHTS=backend/models/best.pt

4. Configure desk zones

Edit backend/config.py → DESK_ZONES. Each entry is {id, label, x1, y1, x2, y2} in pixel space of the camera view. Or PUT zones at runtime via /api/desks.

5. Run the API

bash

python -m backend.scripts.websocket_server
# or
uvicorn backend.app.main:app --host 0.0.0.0 --port 8000

Endpoints:

—GET / service info
—GET /docs OpenAPI / Swagger
—WS /ws live JSON frame results
—GET /api/stream.mjpg annotated MJPEG preview
—GET /api/desks list desks
—PUT /api/desks replace desks
—POST /api/source { "source": "rtsp://..." | "path.mp4" | "0" }
—POST /api/control/stop pause pipeline
—POST /api/control/restart
—GET /api/analytics/summary | series | alerts
—POST /api/analytics/export CSV download
—GET /api/latest last frame result

6. Inference utilities

bash

# Annotate a video to mp4
python -m backend.scripts.video_inference --source path/to/video.mp4 \
    --output backend/output/annotated.mp4

# Live RTSP window
python -m backend.streams.rtsp_stream rtsp://user:pass@192.168.1.10:554/stream1

Training

bash

# 1. extract frames at 2 fps
python -m backend.scripts.extract_frames --input Dataset/videos \
    --output Dataset/raw/images --fps 2

# 2. annotate with Label Studio / Roboflow → see backend/training/annotate_guide.md
#    end with Dataset/raw/images/*.jpg and Dataset/raw/labels/*.txt

# 3. split + dataset.yaml
python -m backend.training.prepare_dataset --input Dataset/raw --output Dataset/yolo

# 4. train YOLOv8l
python -m backend.training.train_yolov8 --data Dataset/yolo/dataset.yaml \
    --epochs 80 --batch 16 --imgsz 640 --device 0

# 5. validate
python -m backend.training.validate_model --weights backend/models/best.pt \
    --data Dataset/yolo/dataset.yaml

Frontend

bash

cd frontend
npm install
npm run dev          # http://localhost:5173

Set the backend URL in frontend/.env (defaults to http://localhost:8000):

VITE_API_URL=http://localhost:8000
VITE_WS_URL=ws://localhost:8000/ws

Notes

—The system is centroid-based: when a detected person's bounding-box center falls inside a desk rectangle, that desk is marked occupied.
—No chair detection, no segmentation, no pose model — by design.
—For best results, fine-tune YOLOv8l on a few thousand frames of your actual camera angles. Generic COCO weights work but degrade on extreme overhead views.