Radar-Based Human Activity Recognition: What Researchers Need to Know Before Starting

A researcher downloads a radar dataset, converts the samples into spectrograms, trains a convolutional neural network, and obtains 97% accuracy. Has the researcher built a reliable human activity recognition system?

Possibly—but the accuracy alone tells us very little.

Were the training and test samples recorded from different people? Did overlapping windows from the same movement appear in both sets? Was the test conducted in another room? What information had already been discarded before the data reached the model? Would the system recognize an activity that was not included in its training labels, or would it confidently force that activity into a known class?

These questions are not secondary details. They determine what the reported performance actually means.

Radar-based human activity recognition, often shortened to radar HAR, lies at the intersection of electromagnetic sensing, signal processing, data engineering, and machine learning. Entering the field through only one of those disciplines can create blind spots. A machine-learning specialist may treat a radar map like an ordinary image. A radar specialist may design excellent range–Doppler processing but underestimate dataset bias. A promising laboratory classifier can fail when the person, viewpoint, room, sensor, or activity vocabulary changes.

This guide develops the decisions that should be made before choosing a neural-network architecture. It is intended for graduate students, researchers, and R&D engineers planning a radar-HAR project.

Central idea: A credible radar-HAR study begins with the sensing problem, measurement geometry, radar configuration, data representation, and evaluation protocol. Model selection comes later.

What is radar-based human activity recognition?

Radar HAR is the process of inferring a human action or motion state from time-varying electromagnetic reflections. A radar illuminates an environment, receives part of the reflected energy, and measures changes associated with the position and movement of scattering points on the body.

The label “human activity recognition” covers several related but distinct problems:

• Presence detection: Is a person present in the monitored area?
• Tracking: Where is the person, and how is that position changing?
• Gesture recognition: Which short, often localized movement is being performed?
• Activity classification: Is the person walking, sitting, standing, bending, falling, or performing another defined action?
• Continuous activity recognition: When does each activity start and end in an uninterrupted stream?
• Pose or keypoint estimation: What is the estimated spatial arrangement of the body or its joints?
• Identification: Which known person produced the observed motion pattern?
• Anomaly or fall detection: Does the observed event deviate from normal behavior or match a safety-relevant event?

These tasks should not be treated as interchangeable. A presence sensor can work with a much simpler observation than a pose-estimation system. A classifier evaluated on pre-cut clips does not solve continuous recognition. A model that separates walking from sitting does not necessarily detect an unknown or abnormal activity.

Defining the exact task is therefore the first research decision.

For a concrete example of a gesture-recognition system—and the limits of its evaluation—see our breakdown of what Google Soli actually proved with 60 GHz radar.

Why use radar for human activity recognition?

The motivation is often summarized in one sentence: radar works without visible light and is more private than a camera. That description is directionally useful but incomplete.

Radar offers several practical advantages:

• It does not form a conventional RGB image of the monitored person.
• Its sensing mechanism does not depend on visible illumination.
• It does not require the subject to wear or charge a device.
• It can measure radial motion directly through the Doppler effect.
• Depending on waveform, bandwidth, and antenna arrangement, it can estimate range, velocity, and direction.
• Small periodic movements can create measurable phase or micro-Doppler patterns.

Those advantages explain its use in research on contactless monitoring, smart environments, assisted living, gesture interfaces, occupancy sensing, and security. Commercial development platforms also make short-range millimetre-wave experiments increasingly accessible. Texas Instruments, for example, documents complete people-counting and tracking pipelines for its IWR6843 family, while Infineon supports radar-based presence-sensing applications. These examples show that the field has an engineering ecosystem beyond academic prototypes. TI people-tracking reference design, Infineon presence sensing

However, radar is not an effortless substitute for a camera.

Human returns may be sparse, aspect-dependent, and mixed with reflections from furniture, walls, fans, doors, or other people. Multipath can create detections at locations where no physical target exists. Similar actions may produce overlapping signatures, while the same action can look different across body types, speeds, directions, distances, and clothing. Interference and hardware settings can also change the data distribution.

The word privacy-preserving deserves particular care. Radar avoids conventional optical imagery, which can reduce some privacy concerns. But motion signatures may still reveal identity, behavior, health-related patterns, location, or daily routines. Radar is better described as a non-optical sensing modality with a different privacy profile, not as automatically anonymous.

First decision: which radar modality fits the task?

“Radar” does not describe one uniform type of data. The waveform and system geometry determine what can be observed.

The table is a starting point, not a universal ranking. The correct modality follows from the desired observable.

If the research question depends mainly on periodic chest displacement, phase-sensitive CW or FMCW processing may be sufficient. If activities must be separated by both their location and velocity, FMCW is usually more appropriate. If the project starts from processed point clouds, it should be recognized that the device firmware has already performed substantial signal processing and thresholding.

What does an FMCW radar actually measure?

FMCW radar is widely used in contemporary radar-HAR research, so it is worth understanding the data before discussing machine learning.

An ideal linear chirp increases its instantaneous frequency with slope $S$S. A target at range $R$R produces a delayed return with round-trip delay

\tau = \frac{2R}{c}

where $c$c is the propagation speed. Mixing the transmitted and received chirps produces a beat frequency. Ignoring Doppler coupling for this simplified explanation,

\begin{aligned}
f_b &\approx S\tau, \\
R &\approx \frac{c f_b}{2S}
\end{aligned}

The ability to distinguish two nearby ranges is governed mainly by the swept bandwidth $B$B:

$$\mathrm{\Delta }R\approx \frac{c}{2B}\mathrm{.}$$

This equation carries an important research lesson: using a higher carrier frequency does not by itself guarantee finer range resolution. Bandwidth is the primary factor in the ideal range-resolution expression.

Movement introduces a Doppler frequency. For a monostatic radar observing radial velocity $v$v,

$$\begin{array}{rl}{\displaystyle f_D}& {\displaystyle =\frac{2v}{\lambda },}\\ {\displaystyle v}& {\displaystyle =\frac{\lambda f_D}{2}\mathrm{.}}\end{array}$$

where $\lambda$λ is the wavelength. Processing samples within a chirp gives range information; processing phase progression across repeated chirps gives Doppler information.

With multiple receive channels, the phase difference across the antenna aperture contains directional information. For an ideal uniform linear array with element spacing $d$d, one simplified relation is

$$\mathrm{\Delta }\varphi =\frac{2\pi d}{\lambda }\sin \theta \mathrm{.}$$

Actual angular resolution depends on aperture, array geometry, signal-to-noise ratio, calibration, estimation method, and the number and placement of targets. Counting antennas alone is not enough.

Texas Instruments provides a useful engineering overview of these range, velocity, and angle dimensions in its millimetre-wave radar fundamentals guide.

Where micro-Doppler enters

A human is not a rigid point target. While the torso moves, the arms, legs, feet, and other body parts have their own time-varying radial velocities. These components modulate the returned signal around the bulk motion and form micro-Doppler structure.

A time–frequency representation can make these patterns visible. Walking may produce periodic limb components around the torso motion; sitting or standing produces a shorter transition; a fall may contain a rapid change followed by low motion. Early work used engineered micro-Doppler features and conventional classifiers such as support vector machines, demonstrating that radar HAR did not begin with deep learning. Kim and Ling, 2009

Micro-Doppler is informative, but it does not make an activity uniquely identifiable. Aspect angle, speed, distance, body characteristics, window length, clutter suppression, and time–frequency parameters all alter the signature.

For a physics-first explanation of how torso and limb motion create these patterns, read Micro-Doppler Explained: How Radar Sees Arms, Legs, and Human Motion. When you are ready to construct the representation, continue with our guide to choosing STFT parameters for radar micro-Doppler.

The complete radar-HAR pipeline

A useful mental model is to separate the pipeline into three domains: sensing, representation, and inference.

Figure 1. A radar-HAR system transforms physical human motion into sampled radar data, processed representations, and finally an activity decision. Original figure created for ArthaVedya.

This sequence matters because errors and information loss propagate forward. A neural network cannot reconstruct an angle dimension that was never measured. It cannot recover weak points removed by a detection threshold unless another input retains them. It may learn room clutter rather than body motion if the dataset permits that shortcut.

A typical FMCW pipeline includes:

1. Acquisition: Samples are collected across ADC time, chirps, receive channels, and frames.
2. Calibration: Channel imbalance, timing, antenna phase, and hardware-specific effects are corrected where required.
3. Interference and clutter handling: Static or slowly varying background components are suppressed according to the task.
4. Range processing: A transform across fast-time samples creates range bins.
5. Doppler processing: A transform across chirps estimates radial-velocity bins.
6. Angle processing: Spatial channels are combined to estimate direction or construct range–angle information.
7. Representation construction: The system retains a radar cube, produces maps, extracts micro-Doppler, or detects points.
8. Temporal segmentation: Frames are grouped into activity examples or processed as a stream.
9. Learning and inference: A model maps the chosen representation to activity labels or other outputs.
10. Decision logic: Smoothing, rejection thresholds, tracking, or event rules convert model outputs into application decisions.

Not every dataset begins at Step 1. Many devices expose only detected points or tracks. That can be ideal for an efficient application, but it limits research on calibration, clutter removal, beamforming, weak-return recovery, and alternative detection methods.

Choose the representation deliberately

The data representation is the bridge between radar physics and machine learning. It decides what information the model can access and what has already been removed.

Figure 2. Comparison of four common radar-HAR representations and the information each preserves or reduces. Original figure created for ArthaVedya.

Representation	Information retained	Information reduced or lost	Suitable research questions
Raw complex ADC/chirp data	Phase and amplitude before most processing	Very little, if metadata and calibration are preserved	End-to-end processing, new clutter suppression, learned beamforming, custom maps
Range profiles over time	Range evolution and reflected magnitude/phase	Doppler and angle may be implicit or collapsed	Range-based transitions, occupancy and motion trajectories
Micro-Doppler spectrogram	Velocity content over time	Often collapses range and angle; STFT choices limit resolution	Gait, gestures, dynamic activity classification
Range–Doppler maps	Joint range and radial velocity	Angle and some phase relationships	Spatially separated motion, detection and activity maps
Range–angle maps	Spatial occupancy and angular structure	Direct velocity detail unless time sequences are used	Pose-like structure, localization, static and slow activities
Radar cube or multi-dimensional map sequence	Multiple range, velocity, angle, channel, and time dimensions	Depends on construction	Rich spatiotemporal learning when compute and data permit
Detected point clouds	Sparse position, Doppler, intensity/SNR attributes	Sub-threshold energy and much raw phase/map structure	Efficient 3D/4D motion learning, tracking, edge deployment
Tracks or engineered features	Compact kinematics or descriptors	Fine scattering structure	Lightweight inference and interpretable baselines

A point cloud is not a lossless radar image

This distinction is frequently missed.

A radar point cloud is normally produced after operations such as FFT processing, beamforming, constant-false-alarm-rate detection, peak selection, coordinate conversion, clustering, and sometimes tracking. Thresholds decide which returns become points. Weak limb reflections may disappear. Multipath may create plausible but false points. Firmware versions and chirp configurations can change the distribution presented to the model.

The 2019 RadHAR project is an influential open example of activity recognition from mmWave point clouds, with voxelized sequences and classical and deep baselines. More recent datasets such as MiliPoint expand the scale and task diversity. MiliPoint’s authors also note that radar chirp settings, despite their influence on detection, are often underreported. RadHAR repository, MiliPoint dataset paper

The correct representation is therefore not simply the tensor that works with a preferred network. It should preserve the variables needed to answer the research question.

Hardware and acquisition choices shape the learning problem

A strong paper should report enough acquisition information for another researcher to understand what the sensor could observe.

Bandwidth and range resolution

Greater swept bandwidth improves ideal range resolution. But usable performance also depends on windowing, sampling, calibration, signal-to-noise ratio, leakage, clutter, and target extent. A person occupies multiple scattering points, so a nominal resolution number is not the same as body-part separability.

Chirp timing and velocity measurement

The interval between chirps, number of chirps, and frame timing affect unambiguous velocity, velocity resolution, and observation duration. A configuration optimized for slow breathing is not necessarily suitable for a rapid gesture or fall.

If the radar platform has not yet been selected, our comparison of 24, 60 and 77 GHz radar for human sensing explains how bandwidth, wavelength, antenna geometry, timing and raw-data access affect the decision.

Frame rate and temporal context

The frame rate must capture the relevant dynamics, while the model window must be long enough to observe the activity. Longer windows add context but also increase latency and may contain multiple activities or transitions.

Antenna aperture and viewpoint

Angle estimation depends on the physical or virtual aperture and its orientation. Sensor height, tilt, azimuth, elevation, and distance determine which body motions project onto radial velocity and which body parts are visible. A movement performed toward the radar can look very different from the same movement performed sideways.

Raw access versus onboard processing

Raw ADC access provides maximum research flexibility but creates larger data volumes and a more demanding processing chain. Onboard point clouds reduce bandwidth and development effort, but bind the project to the manufacturer’s processing assumptions. Neither is universally better; they support different research contributions.

Ground truth and synchronization

Labels entered manually after recording may be sufficient for isolated actions but are inadequate for precise transition timing or keypoint estimation. Cameras, depth sensors, motion capture, wearables, or carefully designed event markers may provide reference data, but they require synchronization and an ethical data-management plan.

A recent 60 GHz dataset for health-oriented motion analysis illustrates the value of richer acquisition reporting: it includes radar point clouds, original radar data, synchronized Kinect-derived skeleton labels, subject identifiers, and daily and rehabilitation-related movements from 26 adults. It should still be interpreted as a controlled healthy-participant dataset, not as clinical proof. Guo, Dogu, and Rajab, 2026

Should you use a public dataset or collect your own?

Public datasets can shorten a project dramatically, provide baselines, and make comparisons possible. But a convenient dataset can quietly redefine the research question.

Before committing to one, ask:

• Does it provide raw ADC data, processed maps, detected points, or tracks?
• Are waveform and antenna parameters documented?
• Are subject, session, environment, viewpoint, and sensor identifiers retained?
• Are recordings continuous, or are activities already segmented?
• Are transitions, background periods, and unknown activities included?
• How many participants and environments are represented?
• Can the official split test the generalization claim you intend to make?
• Were sliding windows created before or after the recordings were divided?
• Are labels temporally precise enough for the intended task?
• Is the license compatible with publication, redistribution, and commercial use?
• Is the dataset still accessible, and are loaders or configuration files available?

The following examples show why dataset names alone are not sufficient:

Sources: RadHAR, MiliPoint, OPERAnet, Spatial mmWave health dataset

Collecting a new dataset is justified when the missing condition—such as a viewpoint, sensor output, population, environment, multi-person scenario, or continuous protocol—is itself central to the research contribution. New data should not be collected merely to avoid understanding an existing benchmark.

Match the model to the representation—not the trend

Once the sensing and representation choices are clear, model selection becomes more disciplined.

Classical baselines

Engineered descriptors with logistic regression, random forests, or support vector machines remain valuable when datasets are small, interpretability matters, or the purpose is to establish whether the representation contains discriminative information. A deep model should outperform a credible simple baseline, not only an intentionally weak one.

Convolutional models

Two-dimensional CNNs are a natural baseline for spectrograms and radar maps, but image-oriented assumptions should be examined. Translation along Doppler does not necessarily mean the same thing as translation along time or range. Resizing a map may alter its physical axes.

Temporal models

Recurrent networks, temporal convolutions, and sequence transformers can model activity evolution. They are useful when the ordering of frames matters, but their context length, latency, and behavior across activity boundaries must be reported. Continuous FMCW classification has been studied explicitly because isolated clips do not capture these difficulties. Shrestha et al., 2020

Point-set and graph models

PointNet-style networks, dynamic graphs, and point transformers can operate on unordered or sequential point sets. They avoid forcing sparse detections into a dense image, but must handle variable point counts, missing weak returns, clutter points, and temporal association.

Large or fashionable models

A transformer is not automatically superior because radar data form a sequence. Model capacity must be supported by dataset size, split difficulty, baselines, ablations, compute constraints, and the deployment target. In many radar-HAR studies, the more meaningful contribution is a representation, protocol, or robustness method rather than another classifier variant.

Evaluation mistakes that can invalidate the result

This is where an apparently excellent project can become scientifically weak.

1. Splitting overlapping windows instead of recordings

Suppose a ten-second recording is converted into four-second windows with a small stride. Neighboring windows share most of their samples. If those windows are randomly divided between training and test sets, the model is evaluated on near-duplicates of its training data.

Figure 3. Overlapping-window leakage occurs when windows from the same recording are distributed across training, validation, and test sets. Independent units should be split before window generation. Original figure created for ArthaVedya.

The safer procedure is:

1. Identify the independent unit: recording, session, participant, or environment.
2. Divide those units into training, validation, and test groups.
3. Generate windows separately inside each group.

This rule also applies when augmentations are derived from the same original recording.

2. Claiming generalization without holding out the relevant factor

Different splits answer different questions:

If the intended application involves new homes, a random within-room split cannot support the deployment claim. Cross-environment adaptation and cross-person generalization remain active research areas precisely because the distribution shift is difficult. Cross-environment radar HAR study, generalizable micro-Doppler point-cloud study

3. Reporting only overall accuracy

Overall accuracy can conceal minority-class failure. At minimum, consider:

• Per-class precision and recall
• Macro F1 or balanced accuracy
• Confusion matrix
• Subject-wise or environment-wise results
• Variability across seeds or folds
• Computational cost and latency

For event-detection systems, frame accuracy may be less meaningful than event precision/recall, detection delay, missed-event rate, and false alarms per hour.

4. Treating the test set as a development set

Repeatedly selecting preprocessing, thresholds, windows, and networks based on test performance leaks information even if no test sample enters training. A validation set is required for those decisions; the test set should support the final evaluation.

5. Comparing models under different pipelines

A model using denoised, range-gated, or manually segmented input is not directly comparable with one operating on raw continuous scenes unless the difference is made explicit. Report the complete pipeline, not only the classifier.

6. Ignoring the negative and unknown world

A closed-set classifier assumes every sample belongs to one of the known classes. A deployed system observes transitions, partial actions, unrelated movements, multiple people, pets, fans, doors, and activities it has never seen. Recent work on open-set micro-Doppler recognition reflects this gap between benchmark classification and realistic inference. Open-set radar HAR study

Laboratory clips are not continuous activity recognition

Pre-segmented clips simplify three problems at once: activity detection, boundary localization, and classification. In a real stream, the system must decide when an event begins, whether a transition belongs to either neighboring class, and how long to wait before making a decision.

A continuous protocol should therefore include:

• Background or null periods
• Natural transitions between activities
• Variable activity duration
• Repeated and interrupted actions
• Temporal annotation rules
• Causal inference if real-time operation is claimed
• Decision latency
• False-alarm and missed-event analysis

Smoothing predictions can reduce flicker but also increase delay and hide short activities. The smoothing rule is part of the evaluated system and should be reported.

Research problems that remain genuinely difficult

Radar HAR is an active field, but novelty should not be claimed simply by replacing a CNN with a newer architecture. More consequential challenges remain.

Cross-person and cross-environment robustness

Body proportions, movement habits, rooms, furniture, sensor placement, and multipath all change the measured distribution. A method that retains accuracy across these shifts is more valuable than a small improvement on a random split.

Viewpoint and range dependence

Radar measures radial motion most directly. A limb moving across the radar’s line of sight may produce less Doppler than the same limb moving toward it. Multi-view datasets and geometry-aware representations can help expose this limitation rather than hiding it.

Open-set recognition

A practical system must sometimes say, “This is not one of the activities I know.” Calibrated uncertainty, rejection mechanisms, and evaluation on meaningful unknown classes remain important.

Multiple people

Simultaneous targets create detection association, occlusion, overlapping micro-Doppler, and identity-maintenance problems. Recent work using sequential 4D point clouds shows growing attention to multi-human activity recognition, but the problem remains much harder than single-person clips. Multi-human FMCW radar HAR study

Label efficiency and realistic data scale

Collecting synchronized, diverse, well-labelled radar data is expensive. Self-supervised learning, simulation, transfer learning, and multimodal supervision are promising only when their domain assumptions and leakage risks are tested carefully.

Transfer across radar configurations

A model can learn the binning, thresholds, noise, and firmware behavior of one device. Research should distinguish human-motion features from sensor-specific signatures and report exactly which settings remain fixed.

Physically meaningful explanations

A heat map over a spectrogram is not automatically an explanation. A stronger analysis connects the model’s evidence to plausible ranges, velocities, body movements, and time intervals, and checks whether that evidence remains stable under nuisance changes.

Edge deployment

Real systems have limits on memory, power, bandwidth, latency, and thermal load. A large offline model may be useful for studying an upper bound, but its deployment relevance should not be assumed.

A practical workflow for starting a radar-HAR project

The following order prevents many avoidable redesigns.

Figure 4. A decision path for selecting radar sensing, data representation, acquisition level, and evaluation protocol before choosing a learning architecture. Original figure created for ArthaVedya.

Step 1: Write the deployment question

Avoid “classify human activities with radar.” A better question is:

Can a wall-mounted radar recognize five daily activities from previously unseen participants in a furnished room while rejecting unrelated movement?

The improved question defines the task, geometry, population, environment, and an unknown/background requirement.

Step 2: Define known, unknown, and transition states

List the target classes and the motions likely to be confused with them. Decide whether the system sees isolated clips or a continuous stream.

Step 3: Identify the physical observables

Ask whether the classes differ mainly through velocity patterns, range trajectory, spatial pose, periodicity, or a combination. This determines whether micro-Doppler, maps, cubes, or point clouds are appropriate.

Step 4: Choose the sensor output level

Decide whether the research requires raw ADC access, processed maps, point clouds, or tracks. Confirm that the hardware and data interface expose that level.

Step 5: Design independent recording units

Plan participants, sessions, viewpoints, rooms, distances, and sensor configurations. Assign the units to splits before creating windows.

Step 6: Establish signal-processing and learning baselines

Build a physically understandable preprocessing baseline and at least one simple classifier. Verify what each additional component contributes through ablation.

Step 7: Test one meaningful generalization axis

A focused project that rigorously studies unseen participants or unseen rooms is more credible than one claiming universal robustness from a random split.

Step 8: Evaluate the complete decision system

Include segmentation, smoothing, confidence thresholds, unknown rejection, runtime, and failure cases when those components are part of the intended use.

Step 9: Record enough detail for reproduction

Preserve radar configuration files, sensor geometry, preprocessing parameters, split identifiers, code versions, and random seeds. Report which data can legally and ethically be shared.

Pre-research checklist

Before collecting data or training a model, confirm the following.

Research question

• The task is defined more precisely than “radar HAR.”
• Target, background, transition, and unknown states are specified.
• The claimed deployment population and environment are stated.

Radar and geometry

• The required observables—range, velocity, angle, phase, or micro-Doppler—are identified.
• Bandwidth, chirp timing, frame rate, antenna geometry, mounting, and range match the task.
• Raw versus processed sensor output is a deliberate choice.

Dataset

• Participants, sessions, rooms, viewpoints, and repetitions are sufficient for the claim.
• Ground truth and synchronization are adequate.
• Background and transition data are included where necessary.
• Ethical approval, consent, anonymization, and licensing are addressed.

Evaluation

• Independent recordings or participants are split before windowing.
• Validation and test sets have separate purposes.
• Metrics reflect class imbalance and event-level behavior.
• At least one realistic distribution shift is evaluated.
• Failure cases, uncertainty, latency, and compute are reported when relevant.

Reproducibility

• Radar configuration and geometry are documented.
• Preprocessing and segmentation parameters are recorded.
• Split identifiers and code versions are preserved.
• Simple baselines and component ablations are included.

Final perspective

Radar-based human activity recognition is appealing because it connects physical sensing with modern inference. That same combination makes the field easy to oversimplify. A radar map is not just another image, a point cloud is not raw radar data, and a high closed-set accuracy is not proof of deployment readiness.

The most important decisions occur before model training:

• What activity problem is being solved?
• Which physical variables distinguish the activities?
• What information does the chosen radar configuration preserve?
• What information does the representation discard?
• Which independent factors must the model generalize across?
• Does the evaluation reproduce those factors honestly?

Researchers who answer those questions clearly are in a much better position to make a meaningful contribution—even with a relatively simple model. In radar HAR, rigorous sensing and evaluation design usually matter more than adding another layer to the network.

References and further reading

1. Texas Instruments, The Fundamentals of Millimeter-Wave Radar Sensors.
2. Y. Kim and H. Ling, Human Activity Classification Based on Micro-Doppler Signatures Using a Support Vector Machine, 2009.
3. A. D. Singh et al., RadHAR: Human Activity Recognition from Point Clouds Generated through a Millimeter-Wave Radar, 2019.
4. A. Shrestha et al., Continuous Human Activity Classification From FMCW Radar With Bi-LSTM Networks, 2020.
5. M. J. Bocus et al., OPERAnet, a Multimodal Activity Recognition Dataset Acquired from Radio Frequency and Vision-Based Sensors, 2022.
6. H. Cui et al., MiliPoint: A Point Cloud Dataset for mmWave Radar, 2023.
7. X. Yang et al., Generalizable Indoor Human Activity Recognition Method Based on Micro-Doppler Corner Point Cloud and Dynamic Graph Learning, 2024.
8. L. Li et al., Radar-Based Cross-Environment Unsupervised Domain Adaptation for Human Activity Recognition, IEEE BIBM, 2025.
9. IEEE Xplore, Open-Set Human Activity Recognition With Micro-Doppler Signatures and Virtual Prototype Learning, document 11224502.
10. IEEE Xplore, Multi-Human Activity Recognition Based on Sequential 4D Point Clouds Using FMCW Radar, document 11008920.
11. K. Guo, E. Dogu, and K. Z. Rajab, A Spatial Millimetre-Wave Radar Point Cloud Dataset for Health Applications, 2026.