⚡ electrical engineering

Lightweight Range-Angle Imaging Based Algorithm for Quasi-Static Human Detection on Low-Cost FMCW Radar

This paper proposes a lightweight, non-visual Range-Angle imaging algorithm using low-cost 60 GHz FMCW radar that significantly outperforms conventional CFAR detectors in accuracy and real-time processing speed for detecting quasi-static human activities in privacy-sensitive indoor environments.

Original authors: Huy Trinh, George Shaker

Published 2026-02-17

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Huy Trinh, George Shaker

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a person sitting quietly on a sofa in a dimly lit, cluttered living room. If you use a standard camera, it might fail because it's too dark, and people might feel uncomfortable being watched. If you use a traditional radar, it's like trying to hear a whisper in a hurricane; the radar is great at spotting fast-moving objects (like a car driving by), but it gets confused by slow, still people because their "signal" is too weak and gets lost in the noise of the walls and furniture.

This paper introduces a clever, low-cost solution to find these "quiet" people using a simple radar and some smart image tricks. Here is how it works, broken down into simple concepts:

1. The Problem: The "Ghost in the Machine"

Standard radar detectors are like security guards who only look for people running. If someone is just lying on a couch, the radar sees them as just another piece of furniture or a wall reflection. It's like trying to find a specific snowflake in a blizzard; the radar gets overwhelmed by the "clutter" (the sofa, the coffee table, the walls) and misses the person.

2. The Tool: A Cheap Radar Eye

The researchers used a very affordable, off-the-shelf radar (the kind you might find in a car's blind-spot monitor) that fits on a small computer board (a Raspberry Pi). They mounted it on a wall to look down at a living room. This radar doesn't take photos; instead, it sends out invisible waves and listens for the echoes.

3. The Magic Trick: Turning Radar Data into a "Heat Map"

Instead of looking at the raw radar data (which is just a bunch of numbers), the team turned it into a 2D picture, similar to a weather map showing temperature.

The Old Way: They tried to use standard math formulas (called CFAR) to find the person. Think of this like trying to find a needle in a haystack by measuring the weight of every single piece of hay. It's slow, and it often misses the needle if the haystack is messy.
The New Way: They treated the radar data like a black-and-white photo. In this photo, the person shows up as a bright, glowing blob, while the walls and furniture are just a dull, gray background.

4. The Solution: The "Top 1%" Filter

The team invented a simple, lightweight algorithm that acts like a smart sieve.

Imagine you have a bucket of mixed sand and gold nuggets. The old methods try to weigh every grain of sand to find the gold.
The new method simply says: "Ignore everything that isn't in the top 1% of brightness."
It looks at the "photo," keeps only the brightest spots (the person), and throws away the dull gray stuff (the clutter).
Then, it uses a simple "connect-the-dots" rule. If the bright spots are close together, it connects them into a single shape. If the shape is too tiny (like a speck of dust), it ignores it. If it's big enough, it says, "There's a person!"

5. The Results: Fast, Accurate, and Private

The results were impressive:

Accuracy: The old radar methods only found the person about 50% to 70% of the time. The new method found them over 93% of the time, even when the person was lying still on a sofa.
Speed: The old methods were like a snail, taking nearly a full second to process one frame. The new method is a cheetah, processing 122 frames every second. It's 74 times faster!
Privacy: Because this uses radar waves instead of a camera, it doesn't take pictures of people. It just sees "a blob of energy." This is perfect for nursing homes or bathrooms where people want privacy but need to be monitored for safety (like if they fall).

The Big Picture

Think of this technology as giving a low-cost radar a pair of "smart glasses." Instead of getting confused by the messy room, it instantly filters out the noise and highlights the human. It's fast enough to run on a tiny computer, cheap enough to put in every room, and private enough to make people feel safe. It turns a difficult problem (finding a still person in a noisy room) into a simple game of "spot the brightest spot."

1. Problem Statement

The paper addresses the challenge of detecting quasi-static human activities (e.g., lying, sitting, standing) in cluttered indoor environments, specifically within Long-Term Care (LTC) facilities.

Detection Difficulty: Unlike moving targets, quasi-static activities produce very low Doppler shifts and highly spread radar signatures. Conventional Constant False Alarm Rate (CFAR) detectors, typically tuned for point targets, struggle to distinguish these weak, diffuse signals from background clutter (walls, furniture).
Privacy and Environment: Cameras are often unsuitable for LTC due to privacy concerns and poor lighting conditions.
Hardware Limitations: Existing deep learning solutions often require high-end MIMO arrays or heavy computational resources, making them unsuitable for low-power, low-cost embedded edge devices.

2. Methodology

The authors propose a lightweight, non-visual, image-based processing pipeline using a low-cost 60 GHz FMCW radar (Infineon XENSIV™ BGT60TR13C) with one transmitter and three receivers.

A. Data Preprocessing & Imaging

Signal Processing: Raw complex IQ data ( $3 \times 128 \times 64$ ) undergoes DC bias removal, windowing (Blackman-Harris), and FFTs over fast and slow time axes to generate Range-Doppler maps.
Clutter Suppression: An exponential moving-average high-pass filter (MTI stage) removes static clutter in the Range-Doppler domain.
Range-Azimuth (RA) Imaging: A Capon beamformer is applied across two spatially separated receive channels (RX0 and RX2) to estimate the spatial covariance matrix. This generates a normalized 2D Range-Azimuth intensity map ( $64 \times 256$ pixels), treating the radar data as a grayscale image.

B. Detection Algorithms

The paper compares three detection methods applied to the RA maps:

2-D CA-CFAR (Cell-Averaging): Standard sliding window approach averaging reference cells to estimate noise power.
2-D OS-CFAR (Order-Statistics): Uses the $k$ -th largest value in the reference window to be more robust against outliers.
Proposed Percentile-Gated Lump Detector:
- Concept: Instead of local thresholding, the algorithm treats the RA map as a global image.
- Step 1 (Gating): Computes a global percentile threshold ( $\tau$ ) (e.g., top 99% of pixel intensities) to create a binary mask, effectively filtering out low-intensity background clutter.
- Step 2 (Morphology): Applies a small binary closing operation ( $2 \times 2$ structuring element) to connect nearby bright pixels and removes tiny blobs below a minimum area ( $A_{min}$ ).
- Step 3 (Selection): Uses a connected-components algorithm to identify blobs and selects the largest connected component as the human target.
- Complexity: $O(HW)$ (linear), significantly faster than the $O(HWK^2)$ complexity of sliding-window CFAR methods.

3. Key Contributions

Novel Approach: First work to propose a lightweight, image-based method specifically tailored for quasi-static human detection using low-cost FMCW radar with limited receive channels.
Algorithm Design: Introduction of a percentile-gated lump detector that leverages global intensity statistics and spatial coherence rather than local noise estimation, making it superior for spread-out, low-velocity targets.
Edge Deployment: Demonstrates that complex radar imaging can be performed on resource-constrained hardware (Raspberry Pi 4B) without deep learning, achieving real-time performance.
Comprehensive Benchmarking: Systematic evaluation against CA-CFAR and OS-CFAR across five semi-static activities and three different subjects in a realistic, high-clutter living room environment.

4. Results

The method was evaluated on a dataset of three subjects performing five activities (standing, sitting on sofa/floor, lying on sofa/floor).

Detection Accuracy:
- Proposed Method: Achieved average accuracies of 93.24% (Subject 1), 92.3% (Subject 2), and 94.82% (Subject 3).
- Comparison: Significantly outperformed CA-CFAR (avg ~59%) and OS-CFAR (avg ~72%).
- Specific Gains: For the difficult "Lying on Sofa" activity, accuracy improved from 33.0% (CA-CFAR) and 43.9% (OS-CFAR) to 86.6% for Subject 1.
Robustness: In high-clutter scenarios (e.g., Frame 106 in the study), CFAR methods failed due to background noise, while the proposed method successfully isolated the human target.
Computational Performance (Raspberry Pi 4B):
- Latency: The proposed algorithm runs at 8.2 ms per frame.
- Throughput: Achieves >122 FPS.
- Speed-up: Represents a 74× speed-up over OS-CFAR (601.1 ms) and a 11× speed-up over CA-CFAR (91.2 ms).
- Resource Usage: Maintains low RAM usage (~299 MB) and meets the 10 Hz radar frame-rate constraint with no deadline misses.

5. Significance

This paper demonstrates that simple image processing primitives can outperform sophisticated statistical detectors (CFAR) and heavy deep learning models for specific radar applications.

Privacy-Preserving: Offers a viable alternative to cameras for monitoring vulnerable populations in LTC facilities.
Scalability: The low computational footprint and reliance on low-cost hardware make the solution highly scalable for deployment in hospitals and large-scale care centers.
Reliability: Proves that robust detection of non-moving humans is possible in cluttered environments without requiring expensive MIMO arrays or high-power GPUs.