CrossCamReID

Real-time People Counting + Re-Identification for multi-camera edge analytics.

What This App Does

production/app_people_counting.py runs a live analytics pipeline that:

• Detects and tracks people per camera.
• Re-identifies the same person across cameras with global IDs.
• Computes occupancy, entry/exit flow, and dwell time.
• Streams results in real time over WebSocket for dashboards.

Why It Feels Enterprise-Ready

• Cross-camera ID consistency using shared ReID identity memory.
• Config-driven behavior (YAML) for controlled deployments.
• Multiple backend options (onnxruntime, tensorrt, fastreid).
• Built-in health endpoint and session lifecycle handling.
• Designed for benchmark-oriented edge inference pipelines.

Architecture (Abstract View)

flowchart TD
    A[Camera Streams\nRTSP/KVS/File] --> B[Detection + Tracking\nYOLO + ByteTrack/BoT-SORT]
    B --> C[ReID Embedding\nTorso crop + feature vector]
    C --> D[Identity Store\nQdrant similarity search]
    D --> E[Global ID Assignment\nQUALIFY -> ENROLL -> LOCK]
    E --> F[People Counting Engine\nOccupancy + Entry/Exit + Dwell]
    F --> G[WebSocket Payloads\nLive metrics + events + optional frames]

Loading

Core Features

• Real-time people counting per camera and organization.
• Cross-camera re-identification with stable person IDs.
• Occupancy analytics and threshold alert support.
• Dwell-time analytics (including cross-camera lifetime dwell).
• Occlusion-aware embedding control to reduce ID pollution.
• Tracker mode switch: bytetrack or botsort.
• Session-aware WebSocket server with token support.

Get Started

1) Install dependencies

pip install -r requirements.txt

Optional extras:

• pip install PyJWT (strict JWT validation)
• pip install boto3 (kvs:// stream support)
• pip install pynvml (GPU stats in payload)

2) Configure the app

Edit:

• production/config/config.yaml

Key sections to verify:

• cameras
• models
• database.qdrant
• runtime.reid_backend
• runtime.tracker_mode

3) Run the People Counting + ReID server

cd production
uvicorn app_people_counting:app --host 0.0.0.0 --port 8002

4) Connect your client

WebSocket endpoint:

ws://<host>:8002/ws/people_counting/{client_id}?token=<JWT-or-dev-token>

Health check:

GET http://<host>:8002/health

Quick Repo Map

production/
  app_people_counting.py               # Main people-counting + ReID app
  dashboard_people_counting_mock.html  # Sample dashboard UI
  config/config.yaml                   # Runtime configuration
  src/crosscamreid/
    websocket/people_counting_handler.py
    websocket/people_counting_runner.py
    counting/                          # occupancy, dwell, entry/exit logic
    reid/                              # ONNX/TensorRT/FastReID backends

Typical Use Cases

• Retail occupancy intelligence
• Campus/building movement analytics
• Entrance/exit flow monitoring
• Multi-camera people insights dashboards

Notes

• Keep camera credentials and secrets outside git-tracked files.
• Use separate config files for dev, staging, and production.
• For benchmark reporting, fix hardware, camera layout, and config version.