Get Started#

This guide walks you through cloning the repository, downloading AI models, and running the NICU Warmer application.

Prerequisites#

Ensure your system meets the System Requirements before proceeding.

1. Clone the Repository#

Use sparse checkout to download only the NICU Warmer component. If you want to clone a specific release branch, replace main with the desired tag. To learn more on partial cloning, check the Repository Cloning guide.

git clone --filter=blob:none --sparse --branch main \
  https://github.com/open-edge-platform/edge-ai-suites.git
cd edge-ai-suites
git sparse-checkout set health-and-life-sciences-ai-suite/NICU-Warmer
cd health-and-life-sciences-ai-suite/NICU-Warmer

2. Manual Model Staging#

Before running make run, stage the workload models in these locations:

  • Repository root:

    • patient detection model files: .xml and .bin

    • person detection model files: .xml and .bin

    • latch detection model files: .xml and .bin

    • action recognition encoder model files: .xml and .bin

    • action recognition decoder model files: .xml and .bin

  • models_rppg/:

    • rPPG workload source model: .hdf5

make run expects these files to exist.

Example Models#

  1. Patient Detection Model - patient-present*

  2. Person Detection Model - people-present*

  3. Latch Detection Model - latch-detect*

    *Download the above model artifacts (.xml and .bin) from Hugging Face and place them in the appropriate model directory structure. Only these files are required for inference; downloading the remaining repository contents is optional.

  4. RPPG Model - MTTS-CAN

  5. Action Recognition Models -

    • Action Recognition Encoder (.xml)

      wget https://storage.openvinotoolkit.org/repositories/open_model_zoo/temp/action-recognition-0001/action-recognition-0001-encoder/FP16/action-recognition-0001-encoder.xml

    • Action Recognition Encoder (.bin)

      wget https://storage.openvinotoolkit.org/repositories/open_model_zoo/temp/action-recognition-0001/action-recognition-0001-encoder/FP16/action-recognition-0001-encoder.bin

    • Action Recognition Decoder (.xml)

      wget https://storage.openvinotoolkit.org/repositories/open_model_zoo/temp/action-recognition-0001/action-recognition-0001-decoder/FP16/action-recognition-0001-decoder.xml

    • Action Recognition Decoder (.bin)

      wget https://storage.openvinotoolkit.org/repositories/open_model_zoo/temp/action-recognition-0001/action-recognition-0001-decoder/FP16/action-recognition-0001-decoder.bin

Third-Party Content

In the course of using these Intel-provided instruction, users may choose to download content (e.g., models, dataset, etc.) created and distributed by third parties. In doing so, these users acknowledge and agree that they have done so after reviewing background information about the content and agreeing to the license governing the content they select.

Notice: Intel does not create the content and does not warrant its accuracy or quality. By accessing the third-party content, or using materials trained on or with such content, you are indicating your acceptance of the terms associated with that content and warranting that your use complies with the applicable license.

3. Prepare Local Assets#

Run setup to verify local assets and generate the rPPG OpenVINO IR when needed:

make setup

This step:

  • checks the staged model files already present in the repo

  • converts .hdf5 to .{xml,bin}

  • preserves existing local assets on repeated runs

Important: make setup must complete before make run. If docker compose up runs first, Docker creates empty directories for missing bind-mount sources, causing pipeline failures.

4. Run the Application#

Start all services (default mixed-optimized device profile):

make run

By default, make run pulls the prebuilt images from Docker Hub.

To build the images locally instead:

make run REGISTRY=false

To pull a specific release tag:

make run TAG=2026.1.0-rc2

This pulls and starts 5 containers:

Service

Port

Purpose

nicu-backend

5001

Flask API + SSE stream + MQTT subscriber

nicu-ui

3001

React dashboard (nginx reverse proxy)

nicu-dlsps

8080

DL Streamer Pipeline Server (GStreamer)

nicu-mqtt

1883

Eclipse Mosquitto MQTT broker

nicu-metrics-collector

9100

Hardware telemetry (CPU/GPU/NPU/Memory)

Device Profiles#

Select a specific device profile at launch:

make run           # Mixed-optimized (GPU detect, CPU rPPG, NPU action)
make run-cpu       # All workloads on CPU
make run-gpu       # All workloads on GPU
make run-npu       # All workloads on NPU

5. Open the Dashboard#

Navigate to http://localhost:3001 in a browser.

Click Prepare & Run to start the AI pipeline. The system will:

  1. Start the GStreamer pipeline with all 5 models

  2. Process video at ~15 FPS

  3. Stream detections and vitals via MQTT

  4. Display results in real-time on the dashboard

6. Stop the Application#

Click Stop in the dashboard, or from the terminal:

make down

Troubleshooting#

Empty directories instead of model files#

If make run was executed before make setup, Docker may have created empty directories for bind-mount paths. Fix:

make down
sudo rm -rf Warmer_Testbed_YTHD.mp4 model_artifacts models_rppg
make setup
make run

Proxy configuration#

If behind a corporate proxy, set environment variables before running:

export HTTP_PROXY=http://proxy.example.com:port
export HTTPS_PROXY=http://proxy.example.com:port
export http_proxy=$HTTP_PROXY
export https_proxy=$HTTPS_PROXY

The compose file forwards these to all containers automatically.