
What Is CFAR Radar Detection?
CFAR (Constant False Alarm Rate) is an adaptive thresholding technique used in radar receivers to decide whether a range-Doppler cell contains a target or just noise. Instead of applying a fixed voltage threshold—which would flood the display with false alarms when clutter rises or miss targets when noise drops—CFAR estimates the local interference power from neighboring cells and scales the threshold so the probability of false alarm Pfa remains at a designer-chosen value, typically 10−4 to 10−8. Every modern automotive, security, and drone-detection radar relies on some variant of CFAR after the FFT stages. For a broader look at the signal chain that precedes CFAR, see how radar works from pulse to detection.
Why a Fixed Threshold Fails

A fixed threshold assumes the noise power is constant across all range bins and all time. In practice, ground clutter, weather returns, and receiver gain drift change the noise floor by 10–30 dB within a single scan. A threshold set for a quiet environment triggers thousands of false alarms in rain; one set for heavy clutter misses small targets in clear air. CFAR solves this by making the threshold track the local interference level, cell by cell.
The Generic CFAR Architecture
All sliding-window CFAR detectors share the same structure:
- Cell Under Test (CUT) — the single range or range-Doppler bin being evaluated.
- Guard cells — a small number of bins on each side of the CUT that are excluded from the noise estimate, preventing target energy from leaking into it.
- Reference (training) cells — N bins on each side whose power values are combined to estimate the local noise or clutter level Z.
- Threshold multiplier α — a scalar derived from the desired Pfa and the number of reference cells. The detection threshold T = α · Z.
If the CUT power exceeds T, a detection is declared. The key design choices are the variant used to combine reference cells, the number of guard and training cells, and the value of α.
Major CFAR Variants
CA-CFAR (Cell-Averaging)
CA-CFAR averages the power in all 2N reference cells. For N cells per side and an exponential (Swerling-0 or Swerling-1) noise model, the threshold multiplier is:
α = N · (Pfa−1/N − 1)
With N = 16 per side (32 total) and Pfa = 10−6, α ≈ 4.68 (linear scale). CA-CFAR is optimal in homogeneous Gaussian noise but suffers in two situations: target masking (a nearby strong target inflates the average and hides a weaker one) and clutter edges (a sudden rise in clutter on one side biases the average upward). Understanding radar clutter and multipath is essential for choosing the right variant.
GO-CFAR and SO-CFAR
Greatest-Of CFAR (GO-CFAR) computes the average on each side separately and takes the larger value, which protects against clutter edges but increases detection loss by about 1 dB. Smallest-Of CFAR (SO-CFAR) takes the smaller average, preserving sensitivity near edges at the cost of more false alarms on the clutter side.
OS-CFAR (Order-Statistic)
OS-CFAR sorts the 2N reference-cell powers and picks the k-th largest value as the noise estimate. Typical choice: k ≈ 3N/4. This is robust against multiple interfering targets—up to 2N − k of them can sit in the reference window without biasing the estimate. The penalty is higher computational cost (sorting) and about 1–2 dB of CFAR loss relative to CA-CFAR in homogeneous noise. OS-CFAR is the default in many drone detection radar designs where small targets cluster at similar ranges.
Comparison Table
| Variant | Noise Estimate | Strengths | Weaknesses | Typical Use |
|---|---|---|---|---|
| CA-CFAR | Mean of 2N cells | Optimal in homogeneous noise; simple | Target masking; clutter-edge false alarms | Open-sky surveillance |
| GO-CFAR | Max of two half-window means | Good at clutter edges | ~1 dB detection loss; target masking persists | Weather radar, maritime |
| SO-CFAR | Min of two half-window means | Preserves sensitivity near edges | High false alarms at clutter boundary | Rarely used alone |
| OS-CFAR | k-th order statistic | Robust to multiple interferers | Sorting complexity; 1–2 dB CFAR loss | Automotive, drone, security |
Setting Guard and Reference Cells
Guard cells prevent the target’s own sidelobes or range-spread energy from contaminating the noise estimate. A rule of thumb: set the number of guard cells per side equal to the target’s expected extent in bins (range resolution cell width divided into the target’s radial extent), plus one or two extra. For a 77 GHz automotive radar with 4 GHz sweep bandwidth, the range resolution is about 3.75 cm; a passenger car extends roughly 4.5 m, so 120 bins—but most energy is in 2–4 bins after windowing, so 3–4 guard cells per side suffices for point targets.
Reference cells should number at least 16 per side (32 total) for a stable noise estimate. Fewer cells raise the variance of the estimate and widen the gap between actual and designed Pfa. More cells improve the estimate but increase the window’s spatial extent, making it span non-homogeneous clutter regions. The balance depends on the scene: a security perimeter radar scanning a fence line faces a very different clutter profile than an open-field sensor.
CFAR Loss and Its Relationship to SNR
Because the noise estimate itself is noisy (it is computed from a finite sample), the effective detection threshold is higher than it would be with perfect noise knowledge. This penalty is called CFAR loss, and for CA-CFAR with 2N reference cells it is approximately:
LCFAR ≈ 10 · log10(1 + 2/N) dB
With N = 16 (32 cells), LCFAR ≈ 0.5 dB. With N = 8, it rises to about 1.0 dB. This loss must be budgeted in the radar link margin alongside SNR and integration gain. In a tight automotive link budget where the required SNR for Pd = 0.9 at Pfa = 10−6 is already 13.1 dB (Swerling-1, single pulse), an extra 1 dB of CFAR loss can mean the difference between detecting a pedestrian at 50 m or 45 m.
Data Rate After the CFAR Detector

After 2D-FFT processing and CFAR detection, only the cells that exceed the threshold are passed downstream as detection reports. The data rate of this detection list depends on the number of targets, the Pfa, and the total number of range-Doppler cells. For a 77 GHz FMCW radar with 512 range bins × 256 Doppler bins = 131 072 cells per frame at 20 frames/s, the raw cell rate is about 2.6 million cells/s. At Pfa = 10−6, the expected false-alarm count is only ~0.13 per frame—negligible. The detection list (target range, velocity, amplitude, angle) typically fits in a few kilobytes per frame, easily carried over SPI or CAN-FD to the domain controller. To check whether your digital interface meets the Nyquist criterion for the ADC sample stream feeding the FFT, use the calculator below.
For example, a 77 GHz radar with a 40 MHz IF bandwidth needs an ADC sampling rate of at least 80 MSPS (complex) or 100+ MSPS with margin. At 12-bit resolution, the raw ADC data rate is roughly 1.2 Gbit/s—well within LVDS or CSI-2 interface limits but important to verify against the FPGA or radar SoC input specification. Undersampling aliases clutter and targets alike, making CFAR thresholds meaningless.
PCB-Level Factors That Affect CFAR Performance
CFAR operates on digitized data, but the noise floor it must distinguish targets from is set by the analog front end—which lives on the PCB.
- Receiver noise figure (NF): Every 1 dB increase in NF raises the noise floor by 1 dB, reducing effective SNR and forcing either a higher detection threshold or a worse Pd. Careful LNA matching on low-loss laminates like RO3003 (Df = 0.0010 at 77 GHz) keeps NF under 10 dB at the system level. Choosing the right substrate is covered in the radar material selection guide.
- Phase noise and spurs: Oscillator phase noise raises the Doppler-domain noise floor, smearing energy across velocity bins and inflating the CFAR reference cells. A clean LO distribution network with short, impedance-controlled traces is critical.
- Power supply ripple: Supply noise at frequencies within the IF bandwidth appears as discrete spurs, which OS-CFAR handles better than CA-CFAR. Proper decoupling and filtered power planes on the radar PCB reduce this risk.
- Antenna sidelobes: High sidelobes admit clutter from off-axis angles, raising the noise estimate and increasing CFAR loss. Sidelobe levels are directly tied to the array aperture taper and element spacing accuracy on the PCB, especially in MIMO virtual array configurations where phase errors compound.
Validating CFAR in the Lab
Before production, CFAR parameters should be validated with a target simulator that injects calibrated returns at known SNR levels into the receiver. Sweeping the injected SNR while monitoring Pd and Pfa produces the receiver operating characteristic (ROC) curve, which should match the theoretical curve within 1–2 dB. Clutter replay—recording real-world returns and playing them back—tests the algorithm against non-homogeneous scenes. Details on setting up these tests are in the radar test and validation guide.
Implementation Tips
- Start with CA-CFAR for simplicity; switch to OS-CFAR only when multi-target or clutter-edge scenarios demand it.
- Apply CFAR independently in range and Doppler (2D-CFAR) or sequentially (range-CFAR then Doppler-CFAR). Sequential is cheaper; 2D is more robust near clutter ridges.
- Log-domain processing (convert power to dB before averaging) simplifies the multiplier to an additive offset, saving a multiply in fixed-point implementations.
- Always simulate with real or recorded clutter data, not just white Gaussian noise. The regulatory band allocation determines which frequencies your test environment will encounter interference on.
- Document the chosen Pfa, N, guard-cell count, and variant in the radar design specification; these are audit points for automotive functional safety (ISO 26262) and must be traceable.
Frequently Asked Questions
What Pfa should I choose for automotive CFAR radar?
Most 77 GHz automotive radars use Pfa between 10−6 and 10−8. Lower values reduce nuisance detections but require higher SNR for the same Pd. The exact choice depends on the tracker’s ability to reject residual false alarms and the safety integrity level (ASIL) of the function.
Can CFAR detect targets in heavy rain clutter?
Yes, provided the target’s return exceeds the local clutter level by enough margin. CFAR adapts to the elevated noise floor from rain, but because rain clutter can vary rapidly in range, GO-CFAR or OS-CFAR handles the transitions better than CA-CFAR. The key is that the target must still have positive SNR relative to the local clutter, not just relative to thermal noise.
How does CFAR differ from a Neyman–Pearson detector?
A Neyman–Pearson detector assumes the noise statistics are perfectly known and sets a fixed threshold to maximize Pd for a given Pfa. CFAR is the practical, adaptive implementation of this idea when noise statistics are unknown and must be estimated from the data. In homogeneous noise with infinite reference cells, CA-CFAR converges to the Neyman–Pearson optimum. See the Wikipedia article on CFAR for a formal derivation.
Does CFAR work in 2D (range-Doppler) or only 1D?
Both. A 1D CFAR across range bins is simplest and most common. A full 2D CFAR uses a rectangular or cross-shaped window in the range-Doppler map, giving better performance near clutter ridges at the cost of more reference cells and computation. Some implementations run 1D CFAR in range, then 1D CFAR in Doppler on the surviving detections—a good compromise.
What is the computational cost of OS-CFAR versus CA-CFAR?
CA-CFAR requires one addition per new cell entering the window (sliding sum), so it is O(1) per cell. OS-CFAR requires sorting 2N values, which is O(N log N) per cell naively, though efficient sliding-window median algorithms reduce this. On a modern radar SoC like the TI AWR series, the hardware accelerator handles both variants at line rate. On an FPGA, CA-CFAR fits easily; OS-CFAR may need a dedicated sort network. The IEEE classic paper by Rohling (1983) remains the reference for OS-CFAR performance analysis.
Ready for a quotation?
Send your requirements for a material, stackup and manufacturing review. Price and schedule are confirmed for your project.
