Release Notes: NICU Warmer#

Version 1.0.0#

2026

This is the initial release of the application, therefore, it is considered a preview version.

NICU Warmer showcases how a single Intel-powered edge system can simultaneously run multiple AI models for neonatal patient monitoring — object detection, contactless vital signs, and action recognition — all within one integrated GStreamer pipeline and React dashboard.

It proves that heterogeneous workloads — from multi-object detection to heart-rate extraction via rPPG and activity classification — can coexist efficiently on one Intel® Core™ Ultra platform without compromising performance or stability.

New

The initial feature set of the application is now available:

  • Real-time object detection: patient, caretaker, and latch clip presence (GPU)

  • Contactless vital signs via rPPG: heart rate and respiratory rate (CPU)

  • Action recognition with 11 NICU-specific categories (NPU)

  • Motion analysis via frame differencing

  • React dashboard with live video, detection overlays, and vitals charts

  • Hardware telemetry monitoring (CPU, GPU, NPU, memory, power)

  • Runtime device configuration via UI settings or REST API

  • Docker Compose deployment with automatic model download (make setup)

  • Device profile presets (all-CPU, all-GPU, all-NPU, mixed-optimized)

Known issues

  • Video upload is limited to 500 MB per file.

  • rPPG accuracy requires adequate lighting and minimal patient motion during the first 10-15 seconds of warmup.

  • If make run is executed before make setup, Docker creates empty directories for bind-mount paths, causing pipeline failures.

    To work around the issue, run:

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

  • NPU fallback to CPU is not reflected in the device profile API response until the pipeline is restarted.

On this page

This Page