Ferkans — Interactive Telecom Tutor

From Optimal Detection to Adaptive Thresholding

Having formed the range-Doppler map, we must decide which cells contain targets and which are noise. The Neyman-Pearson framework (FSI Ch 01) gives the optimal detector when the noise statistics are known. In practice, the noise level varies with range, clutter, and interference — so the threshold must adapt. This is the role of CFAR (constant false alarm rate) detectors.

For imaging, detection applies per-pixel: each voxel in the reconstructed image is tested against a local noise estimate.

Definition:
Neyman-Pearson Detection in Radar

The binary hypothesis test for a single range-Doppler cell:

$H_0: r = w \qquad (\text{noise only}),$ $H_1: r = \alpha s + w \qquad (\text{target present}),$

where $r$ is the matched filter output, $\alpha$ is the unknown target amplitude, and $w \sim \mathcal{CN}(0, \sigma^2)$ .

The Neyman-Pearson detector compares the test statistic $|r|^2$ to a threshold $\gamma$ :

$|r|^2 \underset{H_0}{\overset{H_1}{\gtrless}} \gamma,$

where $\gamma$ is set to achieve a desired false alarm probability: $P_{\text{fa}} = e^{-\gamma/\sigma^2}$ (for complex Gaussian noise with known variance $\sigma^2$ ).

Definition:
Cell-Averaging CFAR (CA-CFAR)

When the noise variance $\sigma^2$ is unknown and varies with range, the CA-CFAR detector estimates it from neighboring cells.

For a cell under test (CUT) at range bin $m$ , the noise estimate uses $2N_r$ reference cells (excluding $N_g$ guard cells on each side):

$\hat{\sigma^2} = \frac{1}{2N_r}\sum_{i \in \text{ref}} |r_i|^2.$

The detection rule is:

$|r_m|^2 \underset{H_0}{\overset{H_1}{\gtrless}} \alpha_{\text{CFAR}} \cdot \hat{\sigma^2},$

where the threshold multiplier for a desired $P_{\text{fa}}$ is:

$\alpha_{\text{CFAR}} = 2N_r \left(P_{\text{fa}}^{-1/(2N_r)} - 1\right).$

The guard cells prevent target energy from leaking into the noise estimate. The reference cells should contain only noise — if a strong target is in the reference window, the noise estimate is inflated and nearby weak targets are masked.

Definition:
Order-Statistic CFAR (OS-CFAR)

The OS-CFAR detector replaces the mean with an order statistic to be robust against interfering targets in the reference window.

Sort the $2N_r$ reference cell powers: $|r_{(1)}|^2 \leq |r_{(2)}|^2 \leq \cdots \leq |r_{(2N_r)}|^2$ .

Choose the $k$ -th order statistic as the noise estimate:

$\hat{\sigma^2} = |r_{(k)}|^2,$

typically with $k \approx 3N_r/4$ .

The detection rule is $|r_m|^2 > \alpha_{\text{OS}} \cdot |r_{(k)}|^2$ .

OS-CFAR is more robust than CA-CFAR at clutter edges and in multi-target environments, at the cost of slightly higher CFAR loss (reduced $P_d$ for the same $P_{\text{fa}}$ ).

Theorem: Detection Probability for Swerling I Targets

For a Swerling I target (exponentially distributed RCS, constant across pulses within a CPI) with average $\text{SNR}$ per pulse $\bar{\gamma}$ , the single-pulse detection probability is:

$P_d = P_{\text{fa}}^{1/(1 + \bar{\gamma})}.$

For $N_p$ pulses with non-coherent integration (square-law combining), the detection probability increases but lacks a closed-form expression — it is computed via:

$P_d = 1 - F_{\chi^2_{2N_p}}\left(\frac{\gamma}{1 + \bar{\gamma}}\right),$

where $F_{\chi^2_{2N_p}}$ is the CDF of a chi-squared distribution with $2N_p$ degrees of freedom and $\gamma = -\sigma^2 \ln P_{\text{fa}}$ .

Proof

Single-pulse case

Under $H_1$ with Swerling I, $|r|^2 \sim \text{Exp}(\sigma^2(1 + \bar{\gamma}))$ . $P_d = P(|r|^2 > \gamma | H_1) = \exp(-\gamma/[\sigma^2(1+\bar{\gamma})])$ .

Relate threshold to $P_{\text{fa}}$

Under $H_0$ , $|r|^2 \sim \text{Exp}(\sigma^2)$ . $P_{\text{fa}} = \exp(-\gamma/\sigma^2)$ , so $\gamma = -\sigma^2\ln P_{\text{fa}}$ .

Combine

$P_d = \exp(\ln P_{\text{fa}}/(1+\bar{\gamma})) = P_{\text{fa}}^{1/(1+\bar{\gamma})}$ .

Example: CA-CFAR Threshold Computation

A radar uses CA-CFAR with $N_r = 16$ reference cells on each side and $N_g = 3$ guard cells. Compute the threshold multiplier $\alpha_{\text{CFAR}}$ for $P_{\text{fa}} = 10^{-6}$ . What is the CFAR loss compared to the ideal fixed-threshold detector?

Solution

Threshold multiplier

With $2N_r = 32$ total reference cells: $\alpha_{\text{CFAR}} = 32 \left((10^{-6})^{-1/32} - 1\right) = 32(10^{6/32} - 1) = 32(10^{0.1875} - 1) = 32(1.538 - 1) = 17.2$ .

CFAR loss

The CFAR loss is the additional $\text{SNR}$ required to achieve the same $P_d$ as a fixed-threshold detector with known $\sigma^2$ . For $2N_r = 32$ , the CFAR loss is approximately $1.2$ dB — a modest penalty for robustness to unknown noise levels.

CFAR Detection Visualization

Apply CA-CFAR and OS-CFAR to a range profile with targets and clutter. Observe how the adaptive threshold tracks the noise floor and how targets near clutter edges are affected.

Parameters

\log_{10} P_{\text{fa}}

-4

2N_r

(reference cells)32

CFAR variant

Add clutter edge

Definition:
M-out-of-N Binary Integration

Binary integration (also called post-detection integration) is a non-coherent technique: apply CFAR detection to each pulse independently, then declare a target present if at least $M$ out of $N$ pulses produce a detection.

For independent pulses with single-pulse detection probability $P_{d,1}$ and false alarm probability $P_{f,1}$ :

$P_d = \sum_{k=M}^{N} \binom{N}{k} P_{d,1}^k (1-P_{d,1})^{N-k},$ $P_{\text{fa}} = \sum_{k=M}^{N} \binom{N}{k} P_{f,1}^k (1-P_{f,1})^{N-k}.$

Binary integration is suboptimal compared to coherent integration but does not require pulse-to-pulse phase coherence.

Common Mistake: Target Masking in CA-CFAR

Mistake:

Using CA-CFAR when multiple targets may exist in the reference window, assuming that the noise estimate is unbiased.

Correction:

A strong target in the reference window inflates $\hat{\sigma^2}$ , raising the threshold and masking nearby weaker targets. This is the target masking problem.

Mitigations: (1) Use OS-CFAR, which selects an order statistic and is robust to a few outliers. (2) Use censored CA-CFAR, which removes outliers before averaging. (3) Use greatest-of (GO-CFAR) which takes the maximum of left and right reference windows, better handling clutter edges.

Historical Note: The Development of CFAR

1960s-1980s

The CFAR concept was developed in the 1960s as radar systems moved from operator-interpreted displays (where the human eye naturally adapted to varying noise levels) to automatic detection. Finn and Johnson (1968) published the first analysis of cell-averaging CFAR for exponential noise. Rohling (1983) introduced OS-CFAR to handle multi-target environments. The progression mirrors a broader theme in radar: replacing human judgment with principled statistical methods, exactly the path from visual interpretation to algorithmic reconstruction in RF imaging.

CFAR (Constant False Alarm Rate)

A detection scheme that adaptively estimates the noise power from reference cells surrounding the cell under test, maintaining a constant false alarm probability despite varying noise levels.

ROC (Receiver Operating Characteristic)

A curve plotting detection probability $P_d$ against false alarm probability $P_{\text{fa}}$ as the decision threshold varies. Parameterized by $\text{SNR}$ and target model (Swerling type).

Quick Check

For a Swerling I target with $P_{\text{fa}} = 10^{-6}$ and average $\text{SNR} = 15$ dB ( $\bar{\gamma} = 31.6$ ), what is the single-pulse detection probability?

$P_d \approx 0.65$

$P_d \approx 0.90$

$P_d \approx 0.99$

$P_d \approx 0.30$

Correction:

P_d \approx 0.65

$P_d = P_{\text{fa}}^{1/(1+\bar{\gamma})} = (10^{-6})^{1/32.6} = 10^{-0.184} \approx 0.655$ .

Key Takeaway

CFAR detectors maintain constant $P_{\text{fa}}$ by adaptively estimating the noise power from reference cells. CA-CFAR is optimal for homogeneous noise but suffers from target masking; OS-CFAR and GO-CFAR are more robust. For Swerling I targets, $P_d = P_{\text{fa}}^{1/(1+\bar{\gamma})}$ — fluctuating targets require significantly more $\text{SNR}$ than non-fluctuating ones.

CFAR Detector: Adaptive Threshold in Action

A cell-averaging CFAR detector slides along a range profile containing targets embedded in spatially varying clutter. The guard cells, reference cells, and the adaptive threshold are shown evolving in real time. Detections are flagged where the cell under test exceeds the local threshold, illustrating how CFAR maintains constant false-alarm rate despite non-stationary background power.

Detection Theory for Radar