
What Is Radar Signal Processing?
Radar signal processing is the digital extraction of target range, velocity, and angle from the raw intermediate-frequency (IF) signal captured by a radar receiver. In a typical FMCW system the processing chain converts a time-domain beat signal into a two-dimensional Range-Doppler map using successive Fast Fourier Transforms (FFTs), then applies detection algorithms to separate real targets from noise. Every stage—from ADC sampling through windowing and FFT to CFAR detection—has parameters that directly affect range resolution, velocity accuracy, and false-alarm rate.
From Chirp to Beat Signal

An FMCW radar transmits a frequency chirp and mixes the received echo with the transmitted signal to produce a beat tone whose frequency is proportional to target range. If the chirp sweeps a bandwidth B in time Tc, and a target sits at range R, the beat frequency is:
fb = 2 · R · B / (c · Tc)
where c ≈ 3 × 108 m/s. A 77 GHz automotive chirp with B = 4 GHz and Tc = 40 µs produces a beat frequency of about 1.67 MHz per metre of target distance. The ADC must sample this IF signal at a rate that captures the maximum expected beat frequency without aliasing—commonly 10–40 MSPS for short-range automotive sensors and up to 100+ MSPS for long-range systems. For a deeper look at how chirp bandwidth sets range capability, see the discussion of FMCW chirp and beat-frequency fundamentals.
The Two-FFT Processing Chain
Range FFT (First FFT)
Each chirp yields N ADC samples. Applying an FFT to those samples converts the time-domain beat signal into a frequency spectrum where each bin corresponds to a discrete range cell. The range resolution is:
ΔR = c / (2 · B)
With B = 4 GHz, ΔR ≈ 3.75 cm—fine enough to separate a pedestrian’s torso from their legs at close range. The number of FFT points determines how many range bins exist; zero-padding beyond N interpolates the spectrum but does not improve true resolution. A detailed treatment of why bandwidth dominates range resolution is available in the radar range resolution guide.
Doppler FFT (Second FFT)
A frame consists of M chirps transmitted in quick succession. After performing the Range FFT on every chirp, the processor stacks the results into an N × M matrix. Each column (one range bin across all chirps) captures the phase progression caused by target motion. A second FFT along this slow-time dimension extracts the Doppler frequency:
fD = 2 · v / λ
where v is radial velocity and λ is the carrier wavelength (≈ 3.9 mm at 77 GHz). The velocity resolution is governed by the total frame time M · Tc. Increasing M sharpens velocity bins but lengthens the coherent processing interval, introducing motion-blur for maneuvering targets. The trade-offs between velocity resolution and Doppler ambiguity deserve careful attention during waveform design.
The Range-Doppler Map
The output of the two-FFT chain is a 2-D Range-Doppler map: range on one axis, velocity on the other, and magnitude (in dB) at each cell. Targets appear as peaks; clutter and noise fill the floor. This map is the starting point for every downstream task—CFAR thresholding, angle estimation via digital beamforming, and tracking.
Windowing: Suppressing Spectral Leakage
Applying a rectangular window (i.e., no window) to the FFT input produces sinc-shaped sidelobes at about −13 dB, which can mask a weak target next to a strong one. Practical radar processors apply a taper before each FFT:
| Window | Mainlobe Width (relative) | Peak Sidelobe (dB) | Typical Use |
|---|---|---|---|
| Rectangular | 1.0× | −13 | Maximum resolution, low dynamic range |
| Hann | 1.5× | −31 | General purpose |
| Blackman-Harris | 2.0× | −92 | High dynamic range, multi-target |
| Chebyshev (−60 dB) | 1.7× | −60 | Balanced resolution and sidelobe control |
Choosing a window is a direct trade-off: wider mainlobe means coarser effective resolution, while lower sidelobes improve detection of weak targets near strong ones. For automotive 77 GHz radars with 4 GHz bandwidth, a Hann or Chebyshev window is common because the inherent range resolution is already very fine and the slight broadening is acceptable.
From Map to Detection: CFAR and Beyond
Once the Range-Doppler map is formed, a detection algorithm decides which cells contain targets. Constant False Alarm Rate (CFAR) processors estimate the local noise floor around each cell under test and set a threshold that maintains a fixed probability of false alarm. Cell-averaging CFAR (CA-CFAR) works well in homogeneous clutter; ordered-statistic CFAR (OS-CFAR) handles clutter edges and multiple targets more gracefully. The choice of CFAR variant, guard cells, and training cells is a system-level decision that feeds back into frame timing and FFT size.
Angle Estimation: The Third Dimension

With multiple receive antennas, the processor adds a third FFT (or more advanced algorithms like Capon or MUSIC) across the spatial dimension to estimate angle of arrival. The angular resolution is roughly λ / (d · NRX), where d is element spacing and NRX is the number of virtual channels. A 12-element virtual array at 77 GHz with half-wavelength spacing achieves about 9.5° 3-dB beamwidth—adequate for automotive lane detection but insufficient for high-resolution imaging, which drives MIMO arrays to 192 virtual channels or more.
Signal Integrity from ADC to FFT
The entire processing chain assumes the digitized IF signal faithfully represents the analog beat tone. In practice, PCB-level signal integrity issues can degrade the data before the first FFT ever runs:
- Clock jitter: ADC sampling-clock jitter of even 200 fs RMS at 40 MSPS introduces a noise floor that limits dynamic range. Keep clock traces short, impedance-matched, and away from switching regulators.
- Power-supply ripple: A 10 mV ripple on the ADC analog supply can appear as a spurious tone in the Range FFT, mimicking a false target at a fixed range bin.
- Crosstalk on IF lines: Coupling between adjacent IF channels in a MIMO receiver creates ghost targets in the angle dimension. Guard traces, ground stitching, and proper stackup planning on appropriate radar laminates are essential countermeasures.
- ADC input matching: A poorly terminated ADC input reflects energy back into the anti-alias filter, creating passband ripple that modulates the beat signal’s amplitude across chirps and raises the Doppler noise floor.
Many assumptions about what does and does not matter for signal integrity on a radar baseband PCB are surprisingly wrong. Use the tool below to test a few common beliefs.
After running through a few myths, you may find that the one about “ground planes don’t need stitching vias at IF frequencies” is false—even at 5 MHz beat frequencies, return-current management matters because the fast clock edges on the same board contain harmonics well into the GHz range. Revisiting your stackup and via-fencing strategy based on these results can prevent noise-floor degradation that no amount of DSP windowing can fix.
Putting It All Together: A Practical Example
Consider a 77 GHz short-range radar designed for occupant detection inside a vehicle cabin, a scenario where the processing chain must distinguish subtle chest-wall motion (respiration at ~0.3 Hz Doppler) from static seat structures:
- Waveform: 256 chirps per frame, B = 1 GHz, Tc = 60 µs, frame time ≈ 15.4 ms.
- Range FFT: 256-point FFT per chirp → range resolution 15 cm, max unambiguous range ~19 m (more than enough for a cabin). Deeper coverage of maximum unambiguous range explains how PRI sets this limit.
- Doppler FFT: 256-point FFT across chirps → velocity resolution ≈ 0.12 m/s, sufficient to detect breathing motion.
- Window: Hann on both dimensions—sidelobes at −31 dB keep strong reflections from the seat frame from masking the weak chest-wall return.
- CFAR: OS-CFAR with 16 training cells and 4 guard cells, because the cabin contains many closely spaced static objects that would bias CA-CFAR.
This example shows that waveform parameters, FFT sizing, window choice, and detection algorithm are all tightly coupled. Changing one—say, doubling B to 2 GHz for finer range resolution—doubles the IF bandwidth and forces a faster ADC, which in turn demands tighter PCB signal-integrity margins.
Frequently Asked Questions
Why are two FFTs needed instead of one?
The first FFT resolves range (fast-time frequency), while the second resolves velocity (slow-time phase change across chirps). A single FFT can extract only one dimension. The two-FFT approach decouples range and Doppler, producing a 2-D map where targets are localized in both dimensions simultaneously.
Does zero-padding improve radar resolution?
No. Zero-padding interpolates the FFT spectrum, making peaks easier to locate, but true range resolution remains c / (2B) and true velocity resolution remains λ / (2 · M · Tc). Only increasing bandwidth or coherent integration time improves actual resolution.
How does radar signal processing differ between FMCW and pulse-Doppler systems?
FMCW radars use a de-chirp mixer that converts range into a low-frequency beat tone, so the ADC rate is modest. Pulse-Doppler radars sample the full RF bandwidth directly (or after down-conversion to a wide IF), requiring much faster ADCs. Both use Range-Doppler FFT processing, but pulse-Doppler systems must also handle range-ambiguity unfolding when using high PRFs.
What FFT size is typical for automotive radar?
Range FFTs of 256–2048 points and Doppler FFTs of 64–256 points are common in 77 GHz automotive sensors. Larger transforms improve bin granularity but increase latency and memory; the choice depends on the radar’s range class and real-time update-rate requirements.
Can windowing be applied in hardware?
Some radar SoCs (e.g., the TI AWR/IWR family) apply windowing coefficients in the on-chip DSP or hardware accelerator before the FFT, reducing processor load. External FPGAs can also implement window multiplication in the data path ahead of an FFT IP core, as described in various window-function references.
Ready for a quotation?
Send your requirements for a material, stackup and manufacturing review. Price and schedule are confirmed for your project.
