[Paper Review] HOOD: Real-Time Human Presence and Out-of-Distribution Detection Using FMCW Radar

Sabri Mustafa Kahya, Muhammet Sami Yavuz+1 more|arXiv (Cornell University)|Jul 24, 2023

Non-Invasive Vital Sign MonitoringEngineering3 citations

TL;DR

HOOD proposes a real-time, reconstruction-based deep learning method for robust human presence and out-of-distribution (OOD) detection using 60 GHz FMCW radar. By leveraging macro and micro range-Doppler images (RDIs) and a multi-encoder multi-decoder architecture with enhanced E-RESPD preprocessing, HOOD achieves 94.36% average AUROC in detecting humans amid clutter, even during breath-holding or partial occlusion, while simultaneously identifying disturbances as OOD.

ABSTRACT

Detecting human presence indoors with millimeter-wave frequency-modulated continuous-wave (FMCW) radar faces challenges from both moving and stationary clutter. This work proposes a robust and real-time capable human presence and out-of-distribution (OOD) detection method using 60 GHz short-range FMCW radar. HOOD solves the human presence and OOD detection problems simultaneously in a single pipeline. Our solution relies on a reconstruction-based architecture and works with radar macro and micro range-Doppler images (RDIs). HOOD aims to accurately detect the presence of humans in the presence or absence of moving and stationary disturbers. Since HOOD is also an OOD detector, it aims to detect moving or stationary clutters as OOD in humans' absence and predicts the current scene's output as "no presence." HOOD performs well in diverse scenarios, demonstrating its effectiveness across different human activities and situations. On our dataset collected with a 60 GHz short-range FMCW radar, we achieve an average AUROC of 94.36%. Additionally, our extensive evaluations and experiments demonstrate that HOOD outperforms state-of-the-art (SOTA) OOD detection methods in terms of common OOD detection metrics. Importantly, HOOD also perfectly fits on Raspberry Pi 3B+ with an ARM Cortex-A53 CPU, which showcases its versatility across different hardware environments. Videos of our human presence detection experiments are available at: https://muskahya.github.io/HOOD

Motivation & Objective

• To address the challenge of false positives in human presence detection caused by moving and stationary clutters in indoor environments.
• To enable robust detection of human presence even when humans are partially visible, holding their breath, or in complex postures.
• To simultaneously detect out-of-distribution (OOD) disturbances (e.g., fans, toys) when no humans are present, ensuring reliable 'no presence' predictions.
• To develop a real-time, activity-free detection system that generalizes across diverse indoor scenarios without relying on human motion cues.
• To outperform state-of-the-art OOD detection methods in radar-based applications using a reconstruction-based deep learning framework.

Proposed method

• HOOD employs a reconstruction-based deep neural network with a multi-encoder multi-decoder architecture trained on macro and micro range-Doppler images (RDIs) from 60 GHz FMCW radar.
• It integrates an enhanced E-RESPD (E-RESPD) pre-processing technique to detect subtle human movements, including respiration and micro-movements, even during breath-holding.
• The system uses multi-thresholding on reconstruction error scores to classify samples as in-distribution (ID) for humans or OOD for disturbances.
• Training includes mixed-ID/OOD samples (ID + OOD), enabling the model to detect humans even when clutter signals dominate.
• The method processes both static (standing, sitting) and dynamic (walking) human behaviors using a unified pipeline without motion-based assumptions.
• HOOD is designed for real-time inference and can be deployed on microcontrollers for edge autonomy.

Experimental results

Research questions

• RQ1 Can a reconstruction-based deep learning model detect human presence in indoor environments with high accuracy, even when clutter signals mimic human motion?
• RQ2 How well can the system distinguish between humans and out-of-distribution disturbances (e.g., fans, moving objects) when no humans are present?
• RQ3 Can the model maintain high detection performance under challenging conditions such as partial occlusion, breath-holding, or crouched postures?
• RQ4 Does incorporating mixed-ID/OOD training samples improve the model’s ability to detect humans in the presence of dominant clutter?
• RQ5 How does HOOD compare to state-of-the-art OOD detection methods in terms of AUROC and FPR95 on radar-based data?

Key findings

• HOOD achieves an average AUROC of 94.36% on a custom 60 GHz FMCW radar dataset, demonstrating high detection performance across diverse scenarios.
• The system maintains high accuracy in detecting humans during breath-holding and crouched postures, where traditional motion-based methods fail.
• HOOD successfully identifies and classifies moving and stationary disturbances as OOD when no humans are present, even in heavy-cluttered rooms.
• The model generalizes well to partially visible humans, such as those obstructed by tables or at oblique angles, maintaining reliable presence detection.
• HOOD outperforms state-of-the-art OOD detection methods in terms of AUROC and FPR95 on radar-based data, particularly in complex mixed-ID/OOD scenarios.
• Real-time evaluation confirms the system’s feasibility for deployment in smart homes, security, and energy-efficient applications.