Constant False Alarm Rate
How CFAR Adapts the Detection Threshold
A radar receiver decides "target present" when the output of the matched filter in a given range-Doppler cell exceeds a threshold. If that threshold were fixed, it would be calibrated for thermal noise only, and any rise in the background, such as a rain cell, a sea-clutter patch, or a land mass entering the beam, would produce a flood of false detections that swamps the tracker. CFAR solves this by making the threshold a multiple of a locally estimated background level. The detector slides a window across the range cells, holds out the cell under test plus a few guard cells on each side, averages (or rank-orders) the remaining reference cells, and multiplies that estimate by a scaling factor chosen to deliver the desired Pfa.
For the cell-averaging detector in homogeneous, square-law-detected noise, the reference-cell sum follows a gamma distribution, which makes the false alarm probability independent of the true noise power. That property is what keeps the rate "constant": only the design Pfa and the number of reference cells N set the multiplier, not the absolute noise level. The cost is CFAR loss, the additional signal-to-noise ratio needed because the threshold rides on a finite, noisy estimate rather than the true mean. That loss shrinks as N grows but can never reach zero.
Cell-Averaging CFAR Math
T = α × Z, where Z = (1/N) ∑i=1N xi (mean of reference cells)
Scaling multiplier (CA-CFAR, square-law, homogeneous):
α = N × (Pfa−1/N − 1)
Example: N = 16, Pfa = 10−6 → α = 16 × (106/16 − 1) ≈ 21.9 (≈ 13.4 dB)
Where T = threshold, Z = noise-power estimate, xi = reference-cell powers, N = reference-cell count, Pfa = design false alarm probability. As N → ∞, α → −ln(Pfa) and CFAR loss → 0 dB.
CFAR Variant Comparison
| Variant | Threshold Estimate | CFAR Loss (N=16, Pfa=10-6) | Strength | Weakness |
|---|---|---|---|---|
| CA-CFAR | Mean of all N cells | ~1.5 to 2 dB | Lowest loss in uniform noise | Masks targets near clutter edges |
| GO-CFAR | Greater of two half-window means | ~2 to 2.5 dB | Controls Pfa at clutter edges | Worse with multiple targets |
| SO-CFAR | Smaller of two half-window means | ~2.5 to 3 dB | Detects targets in adjacent clutter | False alarms at clutter edges |
| OS-CFAR | k-th order statistic (k ≈ 0.75N) | ~2 to 3 dB | Tolerates several interferers | Sorting cost; higher loss |
Frequently Asked Questions
How is the CFAR threshold multiplier alpha calculated for CA-CFAR?
In homogeneous, square-law-detected noise with N reference cells, the scaling factor is α = N × (Pfa−1/N − 1). For N = 16 and a design Pfa of 10−6, α ≈ 16 × (2.371 − 1) = 21.9, about 13.4 dB above the estimated mean noise power. A larger N drives the multiplier toward the ideal −ln(Pfa), but the wider window is more likely to straddle clutter edges or capture interfering targets.
What is CFAR loss and how large is it?
CFAR loss is the extra signal-to-noise ratio a CFAR detector needs versus an ideal fixed threshold that knows the true noise power, because the threshold rides on a finite, noisy estimate. For CA-CFAR it falls with reference-cell count: roughly 3.5 dB at N = 8, about 1.5 to 2 dB at N = 16, and under 1 dB at N = 32 or more, near a Pfa of 10−6. Robust variants such as OS-CFAR add roughly 0.5 to 1 dB more.
When should you use OS-CFAR or GO/SO-CFAR instead of CA-CFAR?
CA-CFAR is optimal only in uniform noise. At clutter boundaries, greatest-of (GO) CFAR suppresses the false alarms CA-CFAR would create on the noisy side. With closely spaced targets, a strong neighbor inflates the average and masks the test cell, so smallest-of (SO) or ordered-statistic (OS) CFAR is better. OS-CFAR ranks the reference cells and selects the k-th value (often near the 0.75N order statistic), tolerating several interferers for about 0.5 to 1 dB more loss.