Micro-Doppler Explained: How Radar Sees Arms, Legs, and Human Motion

Micro-Doppler is the time-varying Doppler modulation produced when parts of a target move differently from the target as a whole. For a walking person, the torso usually creates a comparatively strong bulk-motion component, while swinging arms, legs and feet create faster periodic components around it. A time–frequency plot can therefore reveal gait rhythm, motion direction and the spread of body-part velocities even when the radar cannot form a camera-like image of the person.

Radar does not literally see an arm or a leg in a conventional micro-Doppler spectrogram. It receives a coherent superposition of reflections from many scattering regions and separates some of their motion through Doppler frequency. Interpreting a bright curve as a specific limb is a model-based inference, not direct anatomical imaging.

Short answer: radar “sees” human articulation because every reflecting body region contributes according to its instantaneous radial velocity. The torso forms the central motion trend; limbs repeatedly move faster and slower than that trend, drawing characteristic time-varying Doppler traces.

What micro-Doppler actually means

A rigid target translating at constant radial speed produces an approximately constant Doppler shift. Real targets are rarely perfectly rigid. A helicopter has rotating blades, a machine can vibrate, a bird flaps its wings, and a person moves the torso, arms, legs and feet along different trajectories. Those additional motions modulate the phase and frequency of the returned signal.

Victor Chen and colleagues formalized this distinction between bulk translation and motion of structures on a target, including vibration, rotation, tumbling and coning. Their measured walking example showed an almost constant body Doppler component together with periodic arm motion above and below it (Chen et al., 2006).

The prefix micro describes motion internal to, or superimposed on, the target’s bulk motion. It does not guarantee a tiny Doppler frequency. A rapidly swinging foot can have a larger radial speed than the torso and may extend far from the central body component.

The physics in one equation

Consider one reflecting region on a human body. Let its instantaneous radial velocity be v_r,i(t), defined here as positive when moving toward the radar. For a coherent monostatic radar, its Doppler frequency is

f_{D,i}(t)=\frac{1}{2\pi}\frac{d\phi_i(t)}{dt}=\frac{2v_{r,i}(t)}{\lambda}=\frac{2f_c v_{r,i}(t)}{c}.

Here, φ_i(t) is the measured phase history, λ is wavelength, f_c is carrier frequency and c is propagation speed. Some radar processors use the opposite Doppler sign convention, so direction should always be checked with a controlled approach–departure measurement.

For an articulated target, the radial velocity of the ith scattering region can be separated conceptually into bulk and internal motion:

v_{r,i}(t)=v_{\mathrm{bulk},r}(t)+v_{\mathrm{micro},i,r}(t).

The torso, upper arms, forearms, thighs, lower legs and feet do not arrive as neatly separated signals. A simplified point-scatterer model writes the received slow-time signal as their coherent sum:

s(t)=\sum_{i=1}^{N}a_i(t)\exp\!\left(j\phi_i(t)\right).

The amplitude a_i(t) changes with body shape, material, polarization, orientation, shadowing and interference among scatterers. This is why a real micro-Doppler signature is a textured energy distribution rather than five perfectly clean curves.

How radar distinguishes motion without resolving the body

Spatial resolution and velocity separation are different measurements. A narrowband continuous-wave radar may have no useful range resolution at all, yet it can still separate reflections that occupy different Doppler frequencies. Conversely, two body regions with the same instantaneous radial velocity can overlap in Doppler even if a wideband radar separates them in range.

Think of the return as a chord rather than a photograph. Several motions contribute at once. Time–frequency analysis shows how the chord’s frequency content changes, but assigning each component to a body part requires knowledge of gait mechanics, radar geometry and often additional range or angle information.

This distinction explains why researchers can classify activities from a low-dimensional Doppler spectrogram without reconstructing a skeleton. It also explains why confident pixel-level claims such as “this is the left ankle” are usually unjustified from a single conventional spectrogram. Specialized work has combined micro-Doppler with range information and kinematic models to decompose limb classes, but that is a harder inference problem (Abdulatif et al., 2017).

How walking becomes a micro-Doppler signature

Suppose a person walks toward a radar. The torso advances at a comparatively steady average speed, so it often produces a strong ridge near the bulk Doppler. The body also rocks and changes speed slightly during each step, so the ridge is not perfectly straight.

• Arms: arm swing alternately adds to and subtracts from the torso’s radial motion. The two arms are approximately out of phase and commonly produce weaker periodic structure around the body component.
• Legs: alternating forward and backward swings generate a wider periodic Doppler spread. The exact curves depend on which segments dominate the reflection.
• Feet: distal segments can briefly reach the largest speeds. Their returns may appear as intermittent outer excursions or flashes rather than continuous bright lines.
• Torso: its larger effective reflecting area often makes it the strongest component, but this is an empirical tendency—not a law for every frequency, polarization and view.

Measured X-band data in the foundational study revealed the body component beneath the leg swings, with forward leg swings producing large peaks and the arm modulation tracing a periodic curve. A later 2.4 GHz study likewise reported the strongest spectrogram component from the torso and periodic arm-and-leg modulation around it (Chen et al., 2006; Kim and Ling, 2009).

Figure 1 makes the mapping explicit. The upper panel contains an intentionally simple radial-velocity model. The lower panel forms a coherent 24 GHz return from the same five point scatterers and applies a short-time Fourier transform. The bright central ridge comes from the stronger torso scatterer; the fainter excursions arise from the faster limbs. It is an explanatory simulation, not a measured person and not a biomechanically complete gait model.

Why limb traces cross the torso line

A leg moving forward relative to the torso can have a total radial speed greater than the body. Half a cycle later, its backward swing subtracts from the body’s approach speed. The same scatterer therefore moves from one side of the torso Doppler to the other. If the backward component is strong enough, its total radial velocity can cross zero even while the person continues walking toward the radar.

Left and right limbs alternate, so a complete gait cycle contains repeated, phase-shifted excursions. This periodic structure—not merely total energy—is one reason micro-Doppler is useful for studying walking.

How to read a human micro-Doppler spectrogram

A spectrogram places time on the horizontal axis, Doppler frequency or equivalent radial velocity on the vertical axis, and signal magnitude or power in colour. Begin with the physics of the axes before interpreting the shape.

The safest language is probabilistic: a feature is consistent with a motion mechanism. Stronger anatomical attribution requires supporting range, angle, polarization, multiple views, a validated kinematic model or reference motion capture.

A numerical example at 24, 60 and 77 GHz

Carrier frequency changes how a given radial velocity maps to Doppler frequency. Consider an illustrative torso speed of 1.2 m/s and a fast distal-limb speed of 3.5 m/s. These are example design values, not universal human-motion limits.

For uniformly sampled complex slow-time data, avoiding basic Doppler aliasing requires

\mathrm{PRF}_{\mathrm{eff}}>2\lvert f_{D,\max}\rvert=\frac{4\lvert v_{r,\max}\rvert}{\lambda}.

The practical sampling rate should exceed that boundary with margin for faster-than-expected motion, filters and nonidealities. In time-division-multiplexed MIMO radar, the relevant rate for one virtual channel may be the aggregate chirp rate divided by the number of interleaved transmitters. A visually plausible spectrogram can therefore have a badly aliased velocity axis if the wrong PRF is used.

Higher carrier frequency gives a larger Doppler shift for the same velocity, but it is not automatically the best radar. Bandwidth, antenna aperture, link budget, timing, phase noise, raw-data access and regulation still matter. See the ArthaVedya comparison of 24, 60 and 77 GHz radar for human sensing for the wider engineering trade-off.

Viewing angle can erase the obvious signature

Radar measures motion projected onto its line of sight. For a velocity vector making an angle θ with that line, the simplified projection is

v_r=v\cos\theta.

A person moving directly toward or away from the sensor has a large bulk radial component. Near broadside, that component approaches zero. Each arm and leg has its own three-dimensional trajectory, however, so the entire signature does not simply shrink by one common cosine factor. Some motions vanish, others remain visible, and their relative strengths change.

This is why a classifier trained only on frontal walking can fail at oblique views. Angle variation changes the measurement itself; it is not merely a rotated version of the same spectrogram.

Micro-Doppler is not a radar waveform

The effect can be observed with several coherent radar architectures. The waveform determines what other information accompanies it.

A narrowband CW sensor can therefore produce rich human micro-Doppler without a range–Doppler map. An FMCW device can provide both, but only if the processing preserves the coherent slow-time sequence. Selecting the waveform should follow the measurement need, not the popularity of a particular spectrogram style.

From raw radar data to a useful spectrogram

1. Preserve complex data. Doppler is encoded in phase evolution. Taking magnitude too early discards signed coherent motion information.
2. Calibrate and correct known impairments. Check I/Q imbalance, channel phase, timing, dropped samples and frequency offsets that can shift or blur the result.
3. Perform range processing when the waveform supports it. For FMCW radar, apply the range FFT within each chirp before selecting the complex slow-time data from the target region.
4. Suppress clutter deliberately. Slow-time mean subtraction can reduce stationary clutter, but aggressive high-pass filtering can also erase near-zero or very slow human motion.
5. Select or track the target region. A fixed range gate can lose energy when a person migrates across bins. Declare whether multiple bins are summed in power, combined coherently or followed along a range track.
6. Apply a two-sided time–frequency transform. The STFT is the standard interpretable baseline. Its window duration controls the central time–Doppler trade-off.
7. Calibrate the axes and scaling. Use the effective slow-time PRF, convert Doppler to velocity with the correct wavelength, state the sign convention, and report whether colour represents magnitude or power.
8. Validate with controlled motion. Record a known approach, departure and stationary scene before interpreting complex human data.

Window length, overlap, FFT length and effective PRF can radically change the display. ArthaVedya’s guide to choosing STFT parameters for radar micro-Doppler develops those choices and provides MATLAB and Python implementations. Texas Instruments’ chirp-programming report also shows how FMCW chirp cycle time and frame duration constrain unambiguous velocity and velocity resolution (Dham, 2020).

What controls the visibility of arms and legs?

Carrier frequency

Shorter wavelength maps the same radial velocity to a larger Doppler frequency, which can separate fine motion from a central component more clearly. The receiver still needs sufficient SNR, phase stability and sampling rate. At 2.4 GHz, Kim and Ling found that arm micro-Doppler was largely buried within leg motion for the difficult distinction between ordinary walking and walking while holding a stick.

Aspect and body orientation

A limb moving mainly perpendicular to the radar contributes little instantaneous radial speed. The body can also shadow a far-side arm or leg. Measurements from different aspects should therefore be treated as different sensing conditions.

Reflectivity, polarization and dynamic range

A strong torso return can mask a weak arm component through spectral leakage or limited receiver dynamic range. Clothing, footwear, carried objects and polarization alter the scattering. Brightness is not a direct measure of body-part size.

Bandwidth and range gating

Bandwidth does not create micro-Doppler, but finer range separation can isolate a person from clutter and retain micro-range structure. It can also reveal that energy at one Doppler frequency occupies several body-related range cells.

Clutter removal and normalization

Background subtraction can expose weak motion, yet it can suppress slow activities and distort energy near zero Doppler. Per-image normalization makes signatures visually consistent but removes absolute amplitude differences. A representation suitable for illustration may not be suitable for quantitative comparison.

Multipath and multiple people

Indoor reflections can create delayed, attenuated or sign-altered copies of motion. With CW radar, returns from several people superpose without range labels. FMCW range and angle processing can help, but two people in the same resolution cell remain a source-separation problem.

What information can micro-Doppler support?

Useful information can come from interpretable measurements rather than from the spectrogram as an undifferentiated image:

• bulk or torso Doppler and its change over time;
• upper and lower Doppler envelopes;
• total velocity spread and its asymmetry around the body component;
• gait or limb-motion period;
• energy distribution, intermittency and cadence stability;
• onset, reversal and duration of transient motion.

In a classic measured study, Kim and Ling extracted six related features—torso frequency, total Doppler bandwidth, Doppler offset, torso-region bandwidth, normalized signal-strength variation and limb period—and used them to classify seven controlled activities from 12 participants. Their subject-separated evaluation reached 91.9% average accuracy, while a split containing examples from every participant reached 92.8%. The result demonstrated useful information in the signature, but it did not establish universal performance across sensors, rooms and populations (Kim and Ling, 2009).

Modern systems also learn features directly from spectrograms, range–Doppler sequences or radar cubes. Deep learning can be effective, but its result depends on representation, training data and evaluation design—not merely the network architecture (Gürbüz and Amin, 2019). The broader ArthaVedya guide to radar-based human activity recognition explains subject-independent splits, data leakage and deployment shift.

What micro-Doppler cannot tell you by itself

• Exact three-dimensional anatomy: a single time–Doppler image lacks enough spatial information for unique joint reconstruction.
• Full velocity: ordinary Doppler measures only the line-of-sight component, not the complete velocity vector.
• A unique activity label: different motions can create overlapping signatures, and the same activity changes with person, view and environment.
• Intent or identity: a periodic gait pattern may support a classifier, but it is not direct evidence of who a person is or why they are moving.
• A clinical conclusion: differences in gait or tremor-related motion require appropriate reference measurements, populations and clinical validation before they support a health claim.
• Robust privacy by default: radar does not produce ordinary RGB video, but motion signatures can still reveal activity and potentially identity-related information. Data governance still matters.

Fine hand gestures make the same point at a smaller scale. Google Soli recognized temporal radio-frequency signatures without forming a detailed hand image. ArthaVedya’s Soli paper breakdown separates that demonstrated result from stronger claims the experiment did not establish.

Eight common micro-Doppler mistakes

1. Treating the spectrogram as a body image. It is a time–frequency energy representation of superposed scattering, not a silhouette.
2. Calling every outer ridge a leg. Hands, feet, carried objects, multipath and aliases can occupy similar Doppler regions.
3. Taking magnitude before Doppler processing. This discards the coherent phase evolution that supports signed velocity.
4. Using the ADC sampling rate as the Doppler sampling rate. Micro-Doppler usually uses pulse-to-pulse or chirp-to-chirp slow time.
5. Ignoring TDM-MIMO interleaving. The effective rate for one channel can be much lower than the aggregate chirp rate.
6. Comparing hertz across carrier frequencies. Convert Doppler to radial velocity before comparing motion measured at different wavelengths.
7. Choosing processing by appearance alone. A sharper plot can result from zero-padding, clipping or normalization without containing more physical information.
8. Randomly splitting overlapping clips. Near-duplicate windows from one recording can leak into training and test sets and exaggerate recognition performance.

A defensible experiment checklist

• State carrier frequency, wavelength, waveform and occupied bandwidth.
• Report pulse or chirp timing and the effective slow-time PRF for the analyzed channel.
• Verify maximum expected limb velocity against the unambiguous Doppler interval.
• Document sensor position, subject trajectory, aspect angle, polarization and range.
• Preserve raw complex data and record every calibration and clutter-removal step.
• Declare target range-bin selection, tracking and multi-bin combination.
• Report STFT window type and duration, overlap, FFT length, scaling and displayed dynamic range.
• Include stationary, approach and departure controls to verify clutter removal and Doppler sign.
• Separate participants, sessions and acquisition runs before generating overlapping machine-learning samples.
• Test new people, views, ranges and rooms if the claimed system must generalize to them.

Key takeaways

• Micro-Doppler is time-varying Doppler caused by motion superimposed on a target’s bulk translation.
• A walking torso commonly creates the strong central trend, while arms, legs and feet create faster periodic excursions.
• A spectrogram separates some body motion by radial velocity, not necessarily by range, angle or anatomy.
• Carrier frequency scales Doppler in hertz; aspect angle, PRF, SNR, scattering and preprocessing determine whether limb structure is visible.
• The strongest interpretations combine time–frequency evidence with range, angle, multiple views, kinematic models or reference measurements.
• For machine learning, correct axes and leakage-free evaluation matter more than a visually impressive spectrogram.

The central idea is simple: radar does not need to draw a person to measure articulation. It needs coherent phase, adequate sampling and a geometry that projects meaningful body motion onto the line of sight. Micro-Doppler turns that motion into a signature—but careful processing and restrained interpretation determine what the signature can actually prove.

References

1. V. C. Chen, F. Li, S.-S. Ho and H. Wechsler, “Micro-Doppler Effect in Radar: Phenomenon, Model, and Simulation Study,” IEEE Transactions on Aerospace and Electronic Systems, vol. 42, no. 1, pp. 2–21, 2006. DOI: 10.1109/TAES.2006.1603402.
2. J. L. Geisheimer, W. S. Marshall and E. Greneker, “A Continuous-Wave (CW) Radar for Gait Analysis,” Proceedings of the 35th Asilomar Conference on Signals, Systems and Computers, pp. 834–838, 2001. DOI: 10.1109/ACSSC.2001.987041.
3. J. L. Geisheimer, E. F. Greneker and W. S. Marshall, “High-Resolution Doppler Model of the Human Gait,” Proceedings of SPIE, vol. 4744, 2002. DOI: 10.1117/12.488286.
4. Y. Kim and H. Ling, “Human Activity Classification Based on Micro-Doppler Signatures Using a Support Vector Machine,” IEEE Transactions on Geoscience and Remote Sensing, vol. 47, no. 5, pp. 1328–1337, 2009. DOI: 10.1109/TGRS.2009.2012849.
5. S. Abdulatif, F. Aziz, B. Kleiner and U. Schneider, “Real-Time Capable Micro-Doppler Signature Decomposition of Walking Human Limbs,” 2017 IEEE Radar Conference, pp. 1093–1098. DOI: 10.1109/RADAR.2017.7944367.
6. S. Z. Gürbüz and M. G. Amin, “Radar-Based Human-Motion Recognition With Deep Learning: Promising Applications for Indoor Monitoring,” IEEE Signal Processing Magazine, vol. 36, no. 4, pp. 16–28, 2019. DOI: 10.1109/MSP.2018.2890128.
7. Y. Balal, N. Balal, Y. Richter and Y. Pinhasi, “Time-Frequency Spectral Signature of Limb Movements and Height Estimation Using Micro-Doppler Millimeter-Wave Radar,” Sensors, vol. 20, no. 17, article 4660, 2020. DOI: 10.3390/s20174660.
8. V. Dham, “Programming Chirp Parameters in TI Radar Devices,” Texas Instruments Application Report SWRA553A, revised 2020. Official PDF.