Industry Research

Appearance

FMCW, MIMO, and Doppler Radar Fundamentals ​

FMCW, MIMO, and Doppler Radar Fundamentals curated visual

Visual: FMCW/MIMO radar pipeline from chirps to beat frequency, range FFT, Doppler FFT, virtual array angle, CFAR, and ghost/multipath artifacts.

Automotive and autonomy radars are active RF sensors. They measure range, angle, radial velocity, and reflected power through a signal-processing chain, not through direct geometric projection. Radar point clouds are sparse and noisy compared with LiDAR, but Doppler and weather robustness make radar a critical perception and localization sensor.

1. FMCW Range Measurement ​

Frequency-modulated continuous-wave (FMCW) radar transmits chirps whose frequency changes linearly over time:

f_tx(t) = f_c + S*t
S = B / T_chirp

The received echo is delayed by:

tau = 2*r / c

Mixing received and transmitted signals gives a beat frequency:

f_b ~= S * tau = 2*S*r/c
r = c*f_b/(2*S)

Range resolution is set by chirp bandwidth:

delta_r = c / (2*B)

This is why 77 GHz radars with GHz-class bandwidth can resolve decimeter-scale range bins even though the carrier wavelength is millimeters.

2. Doppler Velocity ​

Doppler comes from phase change across chirps:

f_d = 2*v_r / lambda
v_r = lambda*f_d/2

In a chirp frame, radar usually performs:

Range FFT   over fast time samples within each chirp
Doppler FFT over slow time chirps in a frame
Angle FFT   over antenna channels

The output is a range-Doppler-angle cube. A detection point carries radial velocity relative to the radar:

v_r = dot(v_target - v_sensor, unit_ray)

Radar does not directly measure tangential velocity. A crossing target can have near-zero Doppler even while moving fast.

Sensor Model Impact ​

TaskWhy the model matters
PerceptionDoppler separates moving objects from static clutter and gives immediate velocity without tracking latency.
SLAMStatic radar detections can estimate ego velocity and constrain localization when LiDAR/camera degrade.
MappingRadar maps should store reflectors with RCS/Doppler stability, not assume dense geometry.
ValidationRadar must be validated by range, angle, Doppler, RCS, weather, multipath, and ego-motion bins.

3. MIMO Angle Estimation ​

Multiple-input multiple-output (MIMO) radar uses several transmit and receive antennas to form a larger virtual array.

N_virtual = N_tx * N_rx

For a far-field target, phase differences across antennas encode angle:

delta_phase = 2*pi*d*sin(theta)/lambda

Angle estimation methods:

MethodStrengthRisk
Angle FFTfast and commonresolution limited by aperture and sidelobes
Beamforming/Caponbetter interference handlingmore compute and covariance sensitivity
MUSIC/ESPRITsuper-resolution in ideal conditionsfragile with multipath, coherent targets, calibration errors

Angular resolution depends on aperture, wavelength, SNR, windowing, and target separation. Elevation requires a vertical aperture or planar array.

TDM-MIMO Doppler Coupling ​

Time-division multiplexed MIMO transmits different TX chirps at different times. Moving targets accumulate phase between TX slots. Angle processing must apply Doppler compensation; otherwise velocity appears as angle bias.

4. Radar Signal-Processing Pipeline ​

Typical processing:

ADC samples
  -> range FFT
  -> clutter/static removal or high-pass filtering
  -> Doppler FFT
  -> CFAR detection on range-Doppler map
  -> angle estimation on selected cells
  -> peak grouping / clustering
  -> tracking and classification
  -> point cloud and object list

Windowing trades resolution for sidelobe suppression:

rectangular: narrow main lobe, high sidelobes
Hann/Hamming: wider main lobe, lower sidelobes
Blackman: stronger sidelobe suppression, lower resolution

For fusion, it matters whether a radar point is a raw detection, clustered detection, or tracker output. Tracker outputs already include temporal model assumptions and should not be fused as independent raw measurements.

5. CFAR Detection ​

Constant false alarm rate (CFAR) adapts the detection threshold to local noise and clutter.

Cell-averaging CFAR:

noise_est = mean(training_cells)
threshold = alpha * noise_est
detect if cell_power > threshold

Guard cells around the test cell prevent target energy from contaminating the noise estimate.

CFAR variants:

VariantUse
CA-CFARhomogeneous noise fields
GO-CFARclutter edges, choose greater side estimate
SO-CFARmultiple targets, choose smaller side estimate
OS-CFARrobust to interfering targets using ordered statistic

CFAR defines what becomes a radar point. Low-RCS pedestrians, cones, and FOD may not cross threshold at long range, while bright metal objects can create many detections and sidelobes.

6. RCS, SNR, and Range Equation ​

Radar cross section (RCS) describes how strongly a target reflects energy back toward the radar. Received power follows the radar range equation:

P_r = P_t * G_t * G_r * lambda^2 * sigma /
      ((4*pi)^3 * R^4 * L)

where sigma is RCS and R is range. The R^4 loss is severe: double the range and received power falls by 16x for a point target.

SNR affects:

  • detection probability
  • range and Doppler peak precision
  • angle estimation variance
  • tracker covariance

RCS is not a stable object class label by itself. It changes with aspect angle, material, wetness, polarization, multipath, and target microstructure. Aircraft, GSE, jet bridges, signs, and wet ground can be strong reflectors.

7. Multipath, Ghosts, and Clutter ​

Radar multipath can create detections at geometrically impossible locations. Common causes:

  • ground bounce
  • building and terminal reflections
  • aircraft fuselage reflections
  • guardrails and fences
  • mutual radar interference
  • sidelobes from bright reflectors

Ghost symptoms:

position inconsistent with camera/LiDAR
Doppler inconsistent with apparent location
appears mirrored across a flat reflector
unstable over small ego-motion changes
high RCS but no physical object

Mitigation:

  • fuse with LiDAR/camera occupancy when available
  • use Doppler consistency and multi-frame tracking
  • model static reflector maps for recurring ghost zones
  • reject detections outside drivable/free-space constraints
  • avoid overconfident radar angle covariance

8. Doppler Ego-Velocity Estimation ​

For a static object, measured radial velocity is caused by ego motion:

v_r_i = -dot(v_ego + omega_ego x p_i, unit_ray_i)

With many static radar points, solve for ego velocity:

min_v sum_i rho( v_r_i + dot(v, unit_ray_i) )

Use robust loss rho because dynamic objects and ghosts violate the static assumption.

Ego-Doppler is valuable when:

  • wheel odometry slips on wet apron or snow
  • GNSS velocity is unavailable or multipath-corrupted
  • LiDAR scan matching is weak in rain/fog
  • camera motion estimation fails in low light

Limitations:

  • Requires enough static detections over diverse azimuths.
  • Radial-only geometry is weak when all detections lie in one direction.
  • Mount yaw error biases lateral velocity.
  • TDM-MIMO and Doppler ambiguity handling must be correct.

9. Noise Models for Fusion ​

Radar detection measurement:

z = [r, azimuth, elevation, v_r, rcs]

Approximate covariance:

Sigma_z = diag(sigma_r^2, sigma_az^2, sigma_el^2, sigma_vr^2)

Starting guidance:

  • Range is usually better than angle.
  • Elevation is often worse than azimuth unless using high-channel 4D radar.
  • Doppler is precise for radial motion but ambiguous if velocity wrapping or compensation is wrong.
  • Inflate covariance for low SNR, low RCS, high sidelobe region, multi-target bins, and suspected multipath.

Cartesian covariance:

Sigma_xyz = J_polar_to_cartesian * Sigma_range_angle *
            J_polar_to_cartesian^T

For object tracking, keep radar measurement covariance separate from tracker state covariance. A vendor object list may already be filtered and correlated over time.

10. Radar-Camera and Radar-LiDAR Fusion ​

Fusion roles:

PairRadar contributionOther sensor contribution
Radar + camerasparse metric depth, velocity, weather robustnesssemantics, classification, lane/sign/light recognition
Radar + LiDARDoppler and adverse-weather continuitydense geometry and object shape
Radar + IMU/wheelego-velocity cross-checkangular rate and short-term propagation
Radar + mapstable reflector landmarksglobal frame and semantic constraints

Practical hooks:

  • Calibrate radar yaw/pitch/roll carefully; angle errors dominate at range.
  • Time-align radar frames with ego motion; Doppler compensation uses the correct acquisition interval.
  • Use radar to validate whether a LiDAR/camera object is moving.
  • In rain/fog, raise radar weight but keep multipath/ghost gating active.

11. Failure Modes ​

Failure modeCauseMitigation
Ghost objectmultipath or sidelobeDoppler/geometry consistency, LiDAR/camera cross-check
Wrong velocity signframe or convention mismatchprojection tests with known approaching/receding targets
Angle biasmount yaw/pitch error, antenna calibrationextrinsic calibration and residual monitoring
Doppler aliasingvelocity exceeds unambiguous rangeambiguity flags, multi-PRI schemes, tracker validation
Missed pedestrian/FODlow RCS or CFAR thresholdmulti-sensor fusion and detection tuning
Static clutter treated as moverswrong ego compensationIMU/wheel/GNSS/radar ego-motion consistency
Overconfident radar factorsparse multipath-heavy pointsrobust kernels and covariance inflation

12. Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim