PaperlayerView exact revision on arXiv

Source-linked AI summary

A Review of Indoor Millimeter Wave Device-based Localization and Device-free Sensing Technologies and Applications

Anish Shastri, Neharika Valecha, Enver Bashirov, Harsh Tataria, Michael Lentmaier, Fredrik Tufvesson, Michele Rossi, Paolo Casari

arXiv:2112.05593v3cs.NIcs.LGeess.SP

TL;DR

Indoor mmWave localization and sensing are promising, but practical deployment depends on accessible devices and algorithms that work with their measurements. This survey reviews the propagation, hardware, and approaches used in device-based localization and device-free sensing, finding that current equipment already supports useful capabilities while broader deployment still needs effort.

  • Problem

    Indoor mmWave localization and sensing have practical potential, but democratizing these tasks for vertical applications remains an outstanding need.

    Source: §I · p. 2, §VII · p. 35

  • Method

    The survey reviews mmWave propagation, hardware, localization and sensing techniques, processing methods, and research directions.

    Source: §I · p. 20, D. Key Processing Techniques · p. 25, E. Main learning techniques · p. 26

  • Results

    Current mmWave equipment, including COTS devices, offers opportunities to incorporate localization into communications, and compact commercial radars support precise device-free localization and sensing.

    Source: §VII · p. 35

  • Takeaways & Limitations

    Further effort is needed to make localization and sensing natively available to vertical applications relying on mmWave connectivity.

    Source: §VII · p. 35

    Commercial-grade devices do not expose full internal signal samples and measurements, and their algorithms typically achieve decimeter-level 3rd-quartile localization errors.

    Source: §VII · p. 35

Abstract

from arXiv · show

The commercial availability of low-cost millimeter wave (mmWave) communication and radar devices is starting to improve the penetration of such technologies in consumer markets, paving the way for large-scale and dense deployments in fifth-generation (5G)-and-beyond as well as 6G networks. At the same time, pervasive mmWave access will enable device localization and device-free sensing with unprecedented accuracy, especially with respect to sub-6 GHz commercial-grade devices. This paper surveys the state of the art in device-based localization and device-free sensing using mmWave communication and radar devices, with a focus on indoor deployments. We first overview key concepts about mmWave signal propagation and system design. Then, we provide a detailed account of approaches and algorithms for localization and sensing enabled by mmWaves. We consider several dimensions in our analysis, including the main objectives, techniques, and performance of each work, whether each research reached some degree of implementation, and which hardware platforms were used for this purpose. We conclude by discussing that better algorithms for consumer-grade devices, data fusion methods for dense deployments, as well as an educated application of machine learning methods are promising, relevant and timely research directions.

I. INTRODUCTION

The survey reviews indoor mmWave localization and sensing, whose short wavelengths and propagation characteristics support precise location and target sensing. It distinguishes its coverage from prior surveys and outlines the paper’s treatment of propagation, algorithms, hardware, and open challenges.

I. INTRODUCTION

  • Short wavelengths, quasi-optical propagation, and sparse angular spectra support fine-grained localization and sensing, while higher path loss motivates large arrays and beamforming.Separating multipath components by angle and time can support angle-based localization and target-motion measurements.Source: §I · pp. 1–2

A. Differences with respect to previous surveys

  • Unlike prior surveys focused mainly on sub-6 GHz or outdoor cellular settings, this survey covers indoor mmWave device-based and device-free localization and sensing.It also identifies unsolved challenges and open research avenues, while leaving theoretical limits, standard geometric algorithms, and broader network integration to related surveys.Source: A. Differences with respect to… · pp. 2–3

B. Outline and organization of the manuscript

  • The paper proceeds from propagation and hardware through device-based localization and device-free sensing, then synthesizes findings and research directions.The core technology sections describe techniques, testbed hardware, and surveyed approaches, with summary tables and section-level findings.Source: B. Outline and organization of the… · p. 3

A. Impact of mmWave frequencies on propagation conditions

  • MmWave propagation varies with frequency and materials: diffraction decreases at high frequencies, while rough surfaces and foliage increase diffuse scattering and reduce penetration.Channel models account for these effects using ray tracing or cluster-based methods; ray tracing requires detailed environmental information about surface materials and roughness.Source: A. Impact of mmWave frequencies on… · p. 3

B. Measurement techniques and results

Channel measurements at mmWave frequencies face costly, technically demanding sounders and limitations in angular resolution. The survey reports distinct outdoor and indoor propagation measurements, including path loss, delay, and angular characteristics.

B. Measurement techniques and results

  • High-frequency channel sounders are costly and face increasing phase-noise and antenna-spacing errors, while common directional measurements often use rotating horns whose angular resolution is limited by beamwidth.Switched arrays and super-resolution methods such as SAGE and RIMAX can improve indoor directional and multipath characterization.Source: B. Measurement techniques and results · pp. 3–4

1) Key outdoor results:

  • Outdoor path-loss coefficients are 1.6–2.1 in LoS and 2.5–5 in NLoS, and path-loss standard deviation rises from 5–10 dB at 30 m to over 20 dB at 200 m.The reported variance is attributed to street location and orientation, rather than shadowing.Source: 1) Key outdoor results · p. 4
  • RMS delay spread varies with frequency, so delay windows may better describe the interval containing part of a power-delay profile’s energy.Outdoor angular spreads are about 10° at the BS for one cluster and 30–70° at the UE; BS estimates are considered more accurate because ray tracers may omit scattering objects.Source: 1) Key outdoor results · pp. 4–5

2) Key indoor results:

  • Indoor path-loss coefficients range from 1.2–2 in LoS to 2–3 in NLoS, while human blockage can cause 10–20 dB attenuation.Indoor measurements also show stronger frequency dependence than outdoors, diffuse multipath energy, and scattering from small objects in NLoS conditions.Source: 2) Key indoor results · p. 5
  • Office delay spreads are usually less than 5 ns in LoS and 10–20 ns in NLoS; at 142 GHz, measured values were 3 ns and 9 ns, respectively.Channels above 100 GHz were sparser and had higher partition loss than at 28 GHz; indoor measurements found more multipath components and clusters than outdoor rotating-horn measurements.Source: 2) Key indoor results · p. 5

C. Models for mmWave channels

MmWave channel models must represent three-dimensional angular structure, multipath clusters, and consistency across time, space, and frequency. The survey highlights dynamic blockage and the requirements imposed by large bandwidths and antenna arrays.

C. Models for mmWave channels

  • MmWave channel models require 3D azimuth and elevation spreads, temporal, spatial, and frequency consistency, and cluster-based multipath modeling.The survey notes that multiple models have been developed for indoor and outdoor scenarios, including 3GPP, ITU-R, METIS, and QuaDRiGa.Source: C. Models for mmWave channels · pp. 5–6, D. Summary · p. 6

1) Static vs. dynamic modeling:

  • Dynamic mmWave channels must account for mobility and blockage, which can alter received power and multipath angles and complicate UE localization relative to blockers.Mobility can require more frequent beam training to update AoD and AoA estimates; existing models have added dynamic support for V2X while many remain fixed-BS models.Source: 1) Static vs. dynamic modeling · p. 6, 3) Spatial consistency and clusters · p. 6

D. Summary

3) Spatial consistency and clusters:

  • Spatial consistency requires channel characteristics to evolve smoothly during mobility and account for correlations between nearby UEs sharing a base station.The antenna-element radiation pattern spans azimuth and elevation, and indoor multipath makes polarization properties relevant.Source: 3) Spatial consistency and clusters · p. 6

5) Large bandwidth and large antenna arrays:

  • Large mmWave arrays require channel models to resolve propagation in both angular and delay domains, including AoA and AoD for multiple array geometries.The model must consider array elements in both azimuth and elevation planes.Source: 5) Large bandwidth and large antenna… · p. 6

4) Polarization:

The section compares indoor mmWave beamforming architectures for localization, emphasizing their resolution–cost trade-offs and the limited evidence for indoor channel models.

4) Polarization:

  • Indoor channel-model measurements remain limited beyond sub-6 GHz, with extending measurements to 28 and 140 GHz; Fig. 4 maps UE array patterns across azimuth/elevation scan angles [81].Indoor channel models also incorporate cluster-based modeling, with measurements ongoing for indoor scenarios.Source: 4) Polarization · p. 7, Figure 4 · p. 7

A. Analog beamforming

  • Analog beamforming steers codebook beams with phase shifters and uses RSSI to estimate arrival direction and delay, but limited phase resolution constrains localization.Commercial off-the-shelf devices offer preprogrammed codebooks and access to RSSI and CSI.Source: A. Analog beamforming · p. 7

III. IMPLICATIONS OF BEAMFORMING ARCHITECTURES FOR MMWAVE LOCALIZATION

  • With equal propagation loss across frequencies for equal physical antenna size, mmWave arrays enable directional beams; larger arrays increase beam penetration distance and coverage.Fig. 5 depicts beamwidth relative to operating frequency and array sizes.Source: §III · p. 7, Figure 5 · p. 8

B. Hybrid beamforming

  • Hybrid beamforming combines RF analog and baseband digital processing; it balances resolution, cost, and power, and often performs as well as fully digital while remaining cost-effective and spectrally efficient.Approaches include codebook-dependent, spatially sparse precoding, antenna selection, and beam selection.Source: B. Hybrid beamforming · pp. 7–8

C. Digital beamforming

  • Digital beamforming can provide centimeter-order resolution and real-time ToA/AoA/Doppler updates, but per-antenna RF chains make massive MIMO costly and commercial systems remain immature.The survey notes that only laboratory results exist for fully digital systems.Source: C. Digital beamforming · p. 8

30 GHz

Indoor mmWave localization uses signal measurements and propagation characteristics to estimate device position, with antenna-array design balancing accuracy and complexity. 3GPP Release 16 supports UE localization through downlink and uplink measurements, though the standardized methods do not use the channel’s full dimensionality or phase information.

D. Performance vs. complexity overview

  • Large arrays improve position accuracy, but digital systems increase complexity while analog systems face beam overlap and calibration limits; hybrid beamforming offers a practical tradeoff.The review describes hybrid beamforming as easier to obtain on COTS devices and almost as effective as fully digital systems.Source: D. Performance vs. complexity overview · p. 8

IV. PROGRESS IN STANDARDIZATION OF CELLULAR MMWAVE SYSTEMS

  • Release 16 supports downlink UE positioning from PRS arrival-time differences and uplink positioning from BS/AP measurements of SRS arrival time, power, and AoA.Multiple RTT measurements can also be combined to estimate UE location.Source: §IV · p. 9
  • Standardized methods do not fully use azimuth, elevation, or phase information needed to estimate multipath components at high resolution.The review notes that this remains an active topic in 3GPP Releases 17 and 18.Source: §IV · p. 9

A. Introduction

  • Localization algorithms use power, timing, and angle measurements to estimate distance or direction, then apply geometric or machine-learning methods to infer location.Channel impulse response and CSI can help separate line-of-sight from non-line-of-sight paths or detect when only non-line-of-sight components arrive.Source: A. Introduction · pp. 9–10
  • MmWave’s quasi-optical propagation makes fine signal measurements more feasible, while higher path loss is offset by directional beamforming from large antenna arrays.Beamforming requires access and training protocols, and directs less power toward secondary multipath components.Source: A. Introduction · p. 10

B. Pros and cons of location-dependent measurements for mmWave localization

The review compares signal features and algorithm designs for mmWave localization, noting that angle-based and hybrid approaches predominate. It also describes emerging machine-learning approaches and how their deployment relates to where computation and measurements reside.

B. Pros and cons of location-dependent measurements for mmWave localization

  • Angle measurements are accessible with large arrays, but without range information they provide only relative coordinates, and multipath can distort estimates.CSI can provide rich channel and angle information, although availability and output formats vary by device.Source: B. Pros and cons of location-dependent… · p. 11
  • RSSI offers simple ranging but is error-prone and sensitive to path-loss-model tuning, while ToF can be accurate but requires tight synchronization and sub-nanosecond ADC sampling for fine range resolution.The review also notes that mmWave bandwidth supports better ToF estimation accuracy.Source: B. Pros and cons of location-dependent… · p. 11
  • Hybrid schemes fuse measurements such as AoA with RSSI or ToF, but added complexity and compounded errors can sometimes make them less accurate than non-hybrid methods.Side information can resolve translation, rotation, and scaling ambiguities in some angle-only algorithms.Source: B. Pros and cons of location-dependent… · pp. 11–12
  • Client-centric algorithms estimate location locally, AP-centric algorithms combine client measurements, and cooperative schemes divide localization work between APs and clients.The review observes that ML approaches are mostly AP-centric or cooperative, attributing this to APs’ access to local-server computing power.Source: B. Pros and cons of location-dependent… · pp. 12–14
  • Most surveyed work exploits mmWave arrays’ fine angle resolution, while hybrid approaches combine angle with time or RSSI measurements and are reported to achieve greater accuracy.Only a few works use baseline RSSI, SNR, and time measurements alone; recent surveyed works also apply ML to RSSI and SNR.Source: B. Pros and cons of location-dependent… · p. 13

C. Evaluation tools for mmWave localization

MmWave localization studies evaluate their methods through hardware experiments or simulations, depending on platform availability and measurement flexibility. The survey describes lab and commercial hardware, standard beam training as a source of location-dependent measurements, and ray tracing as a way to model channels and build radio maps.

1) Experimentation-based performance evaluation:

  • Experimental setups range from lab-grade SDRs with mmWave front ends to commercial routers whose firmware exposes beam-training or raw channel measurements.The survey lists, among others, Zynq 7045- and USRP X310-based SDRs, as well as commercial Talon AD7200 routers.Source: 1) Experimentation-based performance… · p. 14
  • Software-defined platforms remain preferred for experiments because of their versatility and multiple digital receiver chains, while COTS hardware is beginning to appear.Using COTS devices can require firmware or operating-system changes to expose radio-chain measurements.Source: 1) Experimentation-based performance… · p. 14
  • Standard beam training yields location-dependent measurements, and known codebooks can support AoA estimation even though generic beam patterns do not map peak received power one-to-one to AoA or AoD.Signal-processing approaches such as compressive sensing can estimate angles without external hardware.Source: 1) Experimentation-based performance… · pp. 14–15

C. Evaluation tools for mmWave localization

  • Ray tracing is used when hardware is unavailable or inflexible, simulating indoor channel features and producing radio maps that can substitute for costly measurement campaigns.The review names NYURay and S5GChannel as examples; custom MATLAB or Python simulations can also generate synthetic datasets.Source: 2) Simulation-based performance… · p. 16

D. Angle-based algorithms

Angle-based localization studies use geometric methods, mapping approaches, and learned regressors to estimate devices from AoA or ADoA measurements. The surveyed results include sub-meter accuracy in reported indoor scenarios, with performance and assumptions varying by method.

D. Angle-based algorithms

  • Single-anchor triangulation, ADoA, and fingerprinting algorithms [122] achieved sub-meter accuracy with high probability when AoA estimate errors were low.The methods were simulated and experimentally validated on prestandard 60-GHz hardware.Source: D. Angle-based algorithms · p. 16
  • Generalized triangulation and ADoA methods achieved sub-meter accuracy in most indoor environments with multiple antennas, with triangulation slightly better in most scenarios.ADoA was preferred for its independence from orientation and compass bias.Source: D. Angle-based algorithms · p. 16
  • JADE jointly maps the client, physical and virtual APs, and indoor environment without prior environmental information, achieving sub-meter accuracy in 90% of simulated cases.It outperformed the approaches in in almost all scenarios, including when AoA estimates were erroneous.Source: D. Angle-based algorithms · p. 16
  • CLAM estimates AP locations, then the client and environment boundaries, and reports sub-meter device-localization errors in about 90% of cases.It fuses AoA-derived ADoA estimates across measurements from different locations.Source: D. Angle-based algorithms · p. 16
  • A shallow neural network trained on ADoA measurements and imperfect JADE estimates [157] achieved sub-meter client accuracy in about 90% of scenarios, even with large AoA errors.Using JADE estimates avoids explicitly collecting a training dataset.Source: D. Angle-based algorithms · p. 16
  • A mobile-client positioning scheme uses coarse AoD information and adaptive downlink beamforming in an indoor mmWave massive-MISO scenario.The passage describes its two-stage design but reports no performance result.Source: D. Angle-based algorithms · p. 16

E. Channel information-based algorithms

The survey covers CSI/CIR-based methods alongside RSSI, SNR, and ToF approaches, noting both their localization uses and practical data-access constraints. Across the reviewed methods, signal features and fingerprints support varied localization strategies, while access to full channel measurements remains limited.

E. Channel information-based algorithms

  • Tensor decomposition of mmWave uplink training signals estimates delay, AoA, and fading coefficients for device localization and tracking.The estimates are transformed into a higher-dimensional tensor using a canonical polyadic model.Source: E. Channel information-based algorithms · p. 17
  • LoS CSI can be converted to angle information for AP–client association, with sub-meter localization in about 80% of cases.The system was implemented on Talon AD7200 routers without interfering with 802.11ad operations.Source: E. Channel information-based algorithms · p. 17
  • Reflected CIR measurements have been used to detect passive objects and model indoor environments in 2D.The method was evaluated with a purpose-built testbed.Source: E. Channel information-based algorithms · p. 17
  • Full CSI/CIR retrieval requires low-level hardware access, available in only a few experimental firmware versions, limiting current use [18].These measurements can map to angle and time information, making them a prospective direction as richer hardware and firmware emerge.Source: E. Channel information-based algorithms · p. 17

F. RSSI and ToF

  • RSSI trilateration has yielded about 1 m average positioning error, while SNR regression on Talon AD7200 routers achieved sub-meter accuracy in 90% of cases,.Other fingerprinting schemes use transmit-beam indices and RSS measurements or spatial-beam SNR values.Source: F. RSSI and ToF · p. 17
  • Two-way round-trip ToF achieved 3 cm average distance estimation and below 5 cm average positioning error, but fine time measurements are often limited to custom high-end transceivers.Beam-training measurements are also used; fingerprinting approaches are gaining popularity and can outperform conventional techniques.Source: F. RSSI and ToF · p. 17

G. Hybrid approaches

The surveyed hybrid approaches combine signal measurements or geometric cues to improve localization or mapping, particularly where individual measurements may fail. Methods range from fingerprints and map-assisted fusion to filtering and multipath-aware schemes.

G. Hybrid approaches

  • Combining two or more measurement techniques can improve localization or mapping accuracy over standalone methods, especially when some measurements fail in challenging environments.The survey notes that most schemes combine angle information with RSSI or time information, often using geometric algorithms.Source: G. Hybrid approaches · pp. 17, 19
  • RSS and AoA methods achieve decimeter-level accuracy even at low SNR, while RSSI/AoA fingerprinting with weighted K-NN achieved sub-meter average accuracy in an experimental UAV system.The fingerprinting approach used particle filters to track UAV motion in GPS-denied indoor environments.Source: G. Hybrid approaches · p. 18
  • mmRanger reconstructs indoor geometry by combining time-domain RSS sequences, reflected-surface AoA/RSS, and robot odometry, with 16 cm mean reflection-point error and 1.72 m maximum error.The implemented system uses path disentanglement and assembles estimated map fragments.Source: G. Hybrid approaches · p. 18
  • Beam-based AoD/SNR positioning yielded centimeter-level errors under LoS conditions, while mWaveLoc achieved 75 cm median 3D error and sub-meter error in 73% of cases.mWaveLoc uses CIR-derived AoA and ToF on IEEE 802.11ad off-the-shelf devices.Source: G. Hybrid approaches · p. 18
  • Map-assisted ToF and angle fusion produced best- and worst-case mean localization errors of about 12 cm and 39 cm, respectively.The method uses an existing or on-the-fly 3D map and clusters candidate user locations.Source: G. Hybrid approaches · p. 18
  • Pseudo-lateration achieved centimeter-level accuracy within 1.5 m and decimeter accuracy beyond it, while motion-assisted filters and multipath-aware approaches address tracking and NLoS errors,,,,.The surveyed work also includes dual-band AoA-assisted mmWave ranging and device-free object tracking from CIR changes, with 6.5 cm mean accuracy.Source: G. Hybrid approaches · pp. 18–19

H. Summary, highlights, and challenges

The survey finds that geometry-based and learned localization methods offer useful accuracy but remain constrained by measurement quality, floor-plan or training requirements, and practical hardware access. Its discussion highlights stronger performance from hybrid information and calls for more accessible platforms and more transferable learning methods.

H. Summary, highlights, and challenges

  • Most geometry-based methods use angle measurements with RSSI or ToF, but depend on accurate measurements and often an accurate indoor floor plan.The survey reports a preference for geometry-based approaches with supporting signal processing.Source: H. Summary, highlights, and challenges · p. 19
  • COTS angle estimates come from CSI decomposition or beam choices, but broad main lobes and sidelobes can cause angle errors.Time measurements from CIR or packet exchange also require tight synchronization between AP and client.Source: H. Summary, highlights, and challenges · p. 19
  • Learned localization models often depend on their training location, while large data requirements and computational complexity constrain use on COTS or embedded devices.These challenges have contributed to comparatively limited application of ML methods to date.Source: H. Summary, highlights, and challenges · p. 19
  • Adaptive and Kalman filtering are commonly used to mitigate location-estimation errors from imperfect measurements, multipath, fading, and NLoS arrivals.The survey cites particle filters, Kalman filters, and LMS filters among the approaches used.Source: H. Summary, highlights, and challenges · p. 19

I. Discussion and future research directions

  • Some lab-grade mmWave schemes achieve decimeter- or sub-decimeter-level accuracy, while hybrid angle and time/RSSI approaches tend to perform even better, though few are tested operationally with COTS equipment.Angle-based AoA/ADoA methods remain accurate, and AoA can be obtained during beam training.Source: §I · p. 20
  • The surveyed localization evaluations are mostly experiments with real hardware, although simulation remains useful for generating large datasets affordably.Table V summarizes the evaluation methods used across surveyed papers.Source: §I · p. 20, Table V · p. 20
  • The survey calls for low-cost standard-compliant mmWave platforms that expose low-level measurements and support research on multi-AP scenarios and zero-initial-knowledge localization and mapping.The proposed direction includes jointly locating APs and clients and estimating indoor floor plans in a SLAM fashion.Source: §I · p. 20
  • Future ML work should reduce training-data needs, federate results across clients and APs, and transfer across previously unseen environments.The motivation is that current training is lengthy and models remain specific to the training area.Source: §I · p. 20

VI. MMWAVE RADAR-ENABLED DEVICE-FREE LOCALIZATION AND SENSING

The radar survey reviews indoor sensing applications and explains how radar reflections support estimates of target properties. It contrasts pulsed-wave limitations with FMCW measurement principles, resolutions, and commonly used hardware platforms.

A. Introduction

  • Radar systems transmit shaped waves and process their reflections to estimate target distance, angle, velocity, and, to some extent, shape.The survey focuses on near-field indoor or city applications, including object tracking and vital-sign monitoring.Source: A. Introduction · p. 23

B. Pulsed Wave Radar

  • Pulsed-wave radar has limited dynamic range and relatively poor SNR at larger distances, so the survey focuses on FMCW for subsequent applications.Pulsed-wave radar may remain viable for short-range uses such as gesture tracking.Source: B. Pulsed Wave Radar · pp. 23–24

C. Frequency-Modulated Continuous Wave Radar

  • FMCW transmits continuous frequency-swept chirps, providing more emitted signal energy than pulsed-wave radar and making it the preferred technology for many mmWave applications.The received signal is mixed with the transmitted signal and low-pass filtered to obtain an intermediate-frequency component.Source: C. Frequency-Modulated Continuous Wave… · p. 24
  • FMCW range follows the TX/RX time delay, and its range resolution is c/(2B), so greater chirp bandwidth improves range resolution.The range is derived from the phase difference of the transmitted and received chirps.Source: C. Frequency-Modulated Continuous Wave… · p. 24
  • FMCW velocity estimation uses phase differences across chirps, with a second Doppler-FFT separating moving objects and velocity resolution given by λ/(2Tf).Here, Tf is the chirp-frame duration.Source: C. Frequency-Modulated Continuous Wave… · p. 24
  • Radar angle estimation uses phase changes across at least two RX antennas, and the stated angle resolution simplifies to 2/N at broadside with half-wavelength spacing.The field of view depends on the maximum measurable AoA.Source: C. Frequency-Modulated Continuous Wave… · p. 24
  • Distance and angle resolution determine radar-map point density and minimum separation, affecting clustering and final tracking performance.The survey notes DBSCAN as one example of a clustering algorithm used to separate subjects and objects.Source: C. Frequency-Modulated Continuous Wave… · pp. 24–25
  • TI commercial evaluation boards are common radar platforms because raw-data software interfaces are available, while some applications still need more powerful or custom hardware.Figure 13 provides an overview of techniques used for mmWave radar sensing applications.Source: Figure 13 · p. 23, C. Frequency-Modulated Continuous Wave… · p. 25

D. Key Processing Techniques

Modern radar processing turns range, velocity, and angle measurements into map representations and temporal signatures that support object sensing. The section explains micro-Doppler and filtering methods used to detect, identify, and track targets.

D. Key Processing Techniques

  • Range, velocity, and angle processing yields RD, RA, or RDA maps; temporal frame sequences support positioning, tracking, and identification.Source: D. Key Processing Techniques · p. 25
  • Micro-motions modulate returned-signal frequency; translation and rotation contribute to micro-Doppler, whose STFT spectrogram can reveal moving bodies and body-part motion.Source: D. Key Processing Techniques · p. 25
  • The discrete Kalman filter alternates state prediction and measurement-weighted correction; it is optimal least-square error estimation for linear systems with additive Gaussian noise.For nonlinear systems, EKF uses a first-order derivative matrix, while UKF uses sigma points and often runs faster because it avoids derivative calculations.Source: D. Key Processing Techniques · pp. 25–26
  • Radar Kalman filters track target centers of mass through noisy or occluded RD, RA, or RDA frames; linear models usually suit human-motion sampling.Source: D. Key Processing Techniques · p. 26
  • Extended-object tracking in retains similar center-of-mass accuracy to point-reflector Kalman filtering while estimating an ellipse for target extent and orientation.Source: D. Key Processing Techniques · p. 26

E. Main learning techniques

This section reviews how clustering and neural-network techniques process mmWave radar data and their use across human sensing, pose estimation, object-related tasks, and health monitoring. It also notes that many health-monitoring results rely on constrained scenarios, leaving robustness in free-living conditions an open challenge.

E. Main learning techniques

  • DBSCAN groups densely connected radar points, rejects outliers, and offers a fast, unsupervised way to cluster targets without specifying their number in advance.It handles non-spherical clusters, runs in O(n log n), and uses a density threshold and neighborhood radius whose settings can remain effective across many scenarios for given hardware and environments.Source: E. Main learning techniques · p. 26
  • CNNs extract features from noisy mmWave radar maps and clusters for object detection, activity recognition, and identity tracking.Convolution and pooling build feature maps for classification; surveyed radar approaches often apply CNNs to clusters produced by DBSCAN. CNNs are translation-invariant but not rotation-invariant.Source: E. Main learning techniques · p. 27
  • RNNs model temporal correlations across radar-map sequences, while autoencoders learn representative features and can reconstruct cleaner radar observations.LSTM and GRU cells address vanishing or exploding gradients at greater complexity; denoising autoencoders are used for noisy maps, including fall detection, surveillance person detection, and indoor identification.Source: E. Main learning techniques · p. 27
  • GANs generate samples such as dense maps from sparse inputs, ResNets use shortcuts to support deep models, and point-cloud networks provide permutation-invariant features for radar data.ResNets are used for skeletal posture estimation and real-time object detection; PointNet-based networks track and identify people from mmWave radar point clouds,, which are lighter to store and process than dense maps but retain less information.Source: E. Main learning techniques · pp. 27–28

F. Selected Applications

  • Human-sensing studies combine radar signatures, clustering, and learned models for recognition, with reported performance varying by task and target count.Identity accuracy in was 90.69% for one target, 97.96% for two, 95.26% for three, and 98.27% for four; [201] reported 97.8% person-detection accuracy for two subjects and 90% for seven, with activity-recognition rates of 92.9% walking, 71.6% running, 99.8% sitting, and 89.9% waving. Other results include activity accuracies from 64% to 99.59% [204] and movement-pattern accuracies from 82.77% to 95.74%.Source: F. Selected Applications · pp. 28–29
  • Radar studies report 17-joint pose errors of 3.2, 2.7, and 7.5 cm on depth, elevation, and azimuth; synthetic-data gestures reach 95.4% accuracy, while two-radar fusion improves accuracy from 46.9% to 98.6%.Source: F. Selected Applications · p. 29
  • Object-related applications use learned detection, segmentation, and sensing methods for obstacles, concealed objects, and hidden targets.SAF-FCOS achieved 90.0% average precision at IoU ≥ 0.50; multi-level EM improved performance by up to 90.0% over conventional EM for concealed-object detection; and HDC reported average precision of 0.69% at IoU 0.5. Hidden-object sensing with a lens-equipped 80–100 GHz device reported probability of detection up to 100%.Source: F. Selected Applications · p. 30
  • Other applications track writing tools and classify objects using radar-derived features, with performance depending on the task and method.The pen-tracking system [243] located pens with 94% accuracy and error below 10 mm across 90% of the trajectory; handwriting tracking achieved 3 mm to 40 mm accuracy across distances of 20 cm to 3 m and 72% to 82% character recognition. Logistic regression and Naive Bayes classified humans, drones, and cars at 86.9% and 73.9%, respectively.Source: F. Selected Applications · p. 30
  • Directly fitting received in-phase and quadrature components with the Levenberg–Marquardt algorithm avoids phase unwrapping and supports beat-to-beat heart-rate and heartbeat-amplitude estimation.The authors identify avoiding phase unwrapping as important in low-SNR scenarios.Source: F. Selected Applications · p. 31
  • 97% breathing-rate accuracy and 83% sleep-position accuracy were achieved by a multi-person monitoring system using high-resolution AoA, an SVM, and an optimal filter.High-resolution AoA identifies closely located targets; the SVM detects sleep position and the optimal filter estimates breathing rate.Source: F. Selected Applications · p. 31
  • 98.4% accuracy was achieved for vital-sign and sleep monitoring, with breathing- and heart-rate mean errors below 0.5 bpm and 2.5 bpm, respectively, at a 70° incident angle.The system first localizes a person by reflection loss, then processes reflected signal-strength samples to estimate vital signs.Source: F. Selected Applications · p. 31
  • 91.08%–97.89% accuracy across eight poses was achieved by a robot that tracks a person, approaches them, and estimates heart rate.The system uses the Hungarian Algorithm and Kalman filtering for tracking, an IIR filter for heartbeat waveforms, and a neural network for heart-rate prediction.Source: F. Selected Applications · p. 31
  • Reflected multi-channel signal signatures from a SOLI mmWave sensor were analyzed using average power spectral density to sense glucose-related changes in blood dielectric constant,.The analysis applies DFT and FFT to each gated signal vector.Source: F. Selected Applications · p. 31
  • Multi-channel maximal-ratio combining outperformed single-channel heart-rate estimates in most of four scenarios that varied radar location and subject posture.The experiments varied both radar placement and the subject’s posture.Source: F. Selected Applications · p. 31
  • The review says many health-monitoring studies use artificial scenarios with stationary people at known locations, while free-living monitoring must handle movement and multi-user tracking without prior location information.Movement can detrimentally affect breathing- and heart-rate estimation, and further experiments are needed to assess general-setting performance and robustness.Source: F. Selected Applications · pp. 31–32
  • Other reported applications include micrometer-level vibration detection, low-visibility semantic mapping, noise-resistant speech reconstruction, and robot localization using interference from other radars.WaveEar reports a 5.5% word error rate at about 1.5 m; the robot-localization system reports position errors from 14 cm with three radars to 6 cm with ten radars. The review notes that reliability under mobility and other noise sources remains unclear.Source: F. Selected Applications · p. 31

G. Summary

Indoor passive mmWave radar research spans diverse sensing applications, with recent work increasingly relying on neural networks. Important gaps remain in robust vital-sign monitoring, multi-radar systems, and their implementation.

G. Summary

  • Neural networks now dominate recent radar-sensing work, including architectures tailored to radar point clouds and temporal signal correlations.The survey also notes GAN-based feature extraction from radar images and good results where analytical models are unavailable.Source: G. Summary · p. 32
  • DBSCAN separates user data in radar maps, while Kalman filtering tracks human trajectories.The survey identifies these as key algorithms for people tracking and activity monitoring.Source: G. Summary · p. 32
  • Applications include people detection, tracking and identification; vital-sign, gesture, activity and fall detection; vibration, speech, and environmental mapping.Source: G. Summary · p. 32
  • Vital-sign monitoring still needs robust algorithms for free-living conditions, including user mobility and other noise sources.Source: G. Summary · p. 32
  • Multi-radar sensing and implementation remain underdeveloped, with synchronization, signal fusion, calibration, deployment architecture, and experimental data among the open needs.The survey notes that analytical and simulated models often fail to represent all noise sources accurately.Source: G. Summary · p. 32

VII. DISCUSSION AND OPEN RESEARCH DIRECTIONS

The review finds that current mmWave equipment already supports localization in communication processes and compact commercial radars support precise sensing, but broader deployment remains a challenge. It identifies hardware access, scalable systems, radar networks, and machine learning as open directions.

VII. DISCUSSION AND OPEN RESEARCH DIRECTIONS

  • Current mmWave equipment, including COTS devices, offers opportunities to incorporate localization into communication, while compact commercial radars cater for precise device-free sensing.Further effort is needed to make these capabilities natively available to vertical applications relying on mmWave connectivity.Source: §VII · p. 35
  • Commercial-grade devices typically achieve decimeter-level 3rd-quartile localization errors, versus centimeter-level accuracy for fully digital architectures and millimeter-level theoretical predictions.Commercial hardware provides limited access to internal samples and measurements, requiring more complex processing and often limiting performance.Source: §VII · p. 35
  • Scalable algorithms should handle multiple APs or clients, ideally estimate the environment without prior floor-plan knowledge, and support integrated sensing and localization.The review also identifies standard-compliant transceivers exposing CSI as a needed hardware direction; none is yet available on the market.Source: §VII · p. 35
  • Radar research could benefit from 2D arrays, massive MIMO, and co-deployed sensors, while requiring multi-view fusion, self-calibration, and deployment architecture choices.The review also points to camera-radar fusion and modifying COTS communication devices such as forthcoming IEEE 802.11ay for passive sensing.Source: §VII · p. 35
  • ML use remains limited in device-based localization compared with radar sensing; fingerprinting typically requires lengthy measurement campaigns, while radar features suit learning and clustering.More features from multiple digital transceiver architectures could also support ML acceleration of intermediate localization steps.Source: §VII · pp. 35–36

VIII. CONCLUSIONS

The survey reviews indoor mmWave localization methods ranging from timing and RSSI to CSI-based approaches, alongside consumer-grade radar sensing applications. It concludes that better consumer-device algorithms, dense-deployment data fusion, and judicious use of machine learning remain open research directions.

VIII. CONCLUSIONS

  • Indoor mmWave localization methods range from timing and RSSI techniques to approaches that process channel state information.The survey also introduces propagation and communication-chain properties relevant to channel measurements.Source: §VIII · p. 36
  • Consumer-grade radar approaches address human activity recognition, object detection, and health monitoring, while open directions include better algorithms, dense-deployment data fusion, and educated ML use.Source: §VIII · p. 36
Loading 2112.05593v3…