How does CA-CFAR calculate the detection threshold?

CA-CFAR computes the threshold by averaging the power values in N reference cells (typically 16-64 cells) surrounding the cell under test (CUT), excluding a set of guard cells immediately adjacent to the CUT. The threshold is T = alpha * (1/N) * sum(X_i), where alpha is the scaling factor that sets the desired false alarm probability. For a desired Pfa and N reference cells with Rayleigh-distributed noise, alpha = N * (Pfa^(-1/N) - 1). With N=32 reference cells and Pfa = 10^-6, alpha = 32 * (10^(6/32) - 1) = 32 * 0.518 = 16.6 (12.2 dB above the noise estimate). Guard cells (typically 2-4 on each side) prevent target energy from leaking into the noise estimate.

What are the limitations of CA-CFAR?

CA-CFAR assumes homogeneous noise/clutter in the reference cells, which fails in two common scenarios. First, at clutter edges (e.g., land-sea boundary), the reference window straddles two regions with different noise levels, causing either excessive false alarms in the low-clutter region or target masking in the high-clutter region. GO-CFAR (Greatest-Of) addresses this by using the maximum of the leading and trailing reference window averages. Second, when multiple targets appear in the reference window, their energy inflates the noise estimate, raising the threshold and masking weaker nearby targets. OS-CFAR (Ordered Statistic) addresses this by using a ranked sample rather than the mean.

How many reference and guard cells should I use?

More reference cells improve the noise estimate accuracy (reducing CFAR loss) but increase computational cost and require homogeneous clutter over a wider range extent. For Rayleigh noise, 16 reference cells yield approximately 1.5 dB CFAR loss relative to an ideal fixed-threshold detector, while 32 cells reduce this to 0.8 dB and 64 cells to 0.4 dB. Guard cells should be sized to span the expected target extent: for a radar with 1 m range resolution and targets spanning 3-5 m, use 3-5 guard cells on each side. Too few guard cells allow target energy to leak into the reference window, biasing the threshold high. Too many waste reference cells and reduce the noise estimate quality.

What is CA-CFAR? | Cell-Averaging CFAR Detector

Definition & Algorithm

CA-CFAR (Cell-Averaging Constant False Alarm Rate) is the most widely used adaptive detection algorithm in radar signal processing. It maintains a constant probability of false alarm (P_fa) across varying noise and clutter environments by estimating the local noise power from a sliding window of reference cells surrounding the cell under test (CUT). The algorithm computes the mean power of the reference cells, multiplies by a scaling factor α derived from the desired P_fa, and declares a detection if the CUT power exceeds this adaptive threshold.

The reference window consists of two halves: leading cells (ahead of the CUT in range) and trailing cells (behind the CUT). Guard cells between the CUT and reference cells prevent target energy spillover from contaminating the noise estimate. CA-CFAR is optimal for homogeneous environments where noise and clutter statistics are uniform across the reference window. In non-homogeneous conditions (clutter edges, multiple targets), variants like GO-CFAR, SO-CFAR, and OS-CFAR provide better performance at the cost of increased CFAR loss or computational complexity.

Key Formulas

Threshold Multiplier (α):

α = N × (P_fa^-1/N − 1)

N = 32, P_fa = 10^-6: α = 32 × (10^6/32 − 1) = 16.6 (12.2 dB)

Detection Threshold:

T = α × (1/N) × ∑ X_i

Where X_i are reference cell power values

CFAR Loss (relative to ideal detector):

L_CFAR(dB) ≈ 10 log₁₀(1 + 2/N)

N = 32: L_CFAR ≈ 0.27 dB; N = 16: L_CFAR ≈ 0.51 dB

CFAR Variant Comparison

Algorithm	Noise Estimate	Clutter Edge	Multi-Target	CFAR Loss	Complexity
CA-CFAR	Mean of all ref cells	Poor	Poor (masking)	Low (0.3-1.5 dB)	O(N)
GO-CFAR	Max of lead/trail means	Good	Poor	Moderate (1-3 dB)	O(N)
SO-CFAR	Min of lead/trail means	Poor	Good	Low-Moderate	O(N)
OS-CFAR	k-th ranked sample	Moderate	Good	Moderate (1-2 dB)	O(N log N)
CASO-CFAR	Censored avg (trimmed)	Moderate	Good	Low-Moderate	O(N log N)

Practical Application

A 77 GHz automotive radar processes 512 range bins from an FMCW beat frequency FFT. CA-CFAR is applied with N = 32 reference cells (16 leading, 16 trailing), 4 guard cells on each side, and P_fa = 10^-4 (threshold multiplier α = 5.3, or 7.2 dB above the noise estimate). In open highway conditions with uniform clutter, CA-CFAR correctly detects vehicles at ranges up to 200 m with probability of detection P_d > 0.95. However, at an overpass boundary where road surface clutter drops abruptly, the trailing reference cells contain higher clutter power than the leading cells, causing the threshold to elevate and masking a vehicle at the transition. The radar firmware switches to GO-CFAR in range bins near detected clutter edges, using the greater of the two half-window estimates to prevent false alarms while accepting 2 dB of additional detection loss.

Frequently Asked Questions

How does CA-CFAR set the threshold?

Averages N reference cells around the CUT, multiplies by α = N(Pfa^(-1/N) - 1). N=32, Pfa=10^-6: threshold is 12.2 dB above noise estimate. Guard cells (2-4 per side) prevent target leakage into the estimate.

What are CA-CFAR's limitations?

Fails at clutter edges (straddles two noise regions) and near multiple targets (inflated noise estimate masks weaker targets). GO-CFAR handles edges; OS-CFAR handles multi-target. CA-CFAR is optimal only in homogeneous clutter.

How many reference/guard cells?

More ref cells = better noise estimate: 16 cells = 1.5 dB CFAR loss, 32 = 0.8 dB, 64 = 0.4 dB. Guard cells should span target extent (e.g., 3-5 cells for 3-5 m targets at 1 m resolution).

CA-CFAR