CFAR Algorithm

The foundational variant that computes the detection threshold by averaging the power of adjacent reference cells surrounding the Cell Under Test (CUT).

• Mechanism: Sums the outputs of leading and lagging reference windows and scales by a threshold multiplier.
• Optimal Environment: Homogeneous background noise with independent, identically distributed (IID) interference.
• Key Weakness: Severe performance degradation in clutter edges and multiple-target scenarios, where interfering signals in reference cells inflate the threshold and mask nearby targets.

Designed to mitigate the clutter edge problem where the noise floor changes abruptly across the range-Doppler map.

• Mechanism: Separately averages the leading and lagging reference windows, then selects the larger of the two sums to compute the threshold.
• Behavior at Edges: Prevents excessive false alarms when the CUT transitions from a low-noise to a high-clutter region by using the window with the higher power estimate.
• Trade-off: Reduced sensitivity to closely spaced targets compared to CA-CFAR, as a target in one window can still dominate the threshold selection.

Optimized for resolving closely spaced targets that would otherwise mutually mask each other in a CA-CFAR detector.

• Mechanism: Computes the average of the leading and lagging reference windows independently, then selects the smaller of the two sums.
• Multi-Target Resolution: If one reference window contains an interfering target, the algorithm relies on the cleaner window to set a lower threshold, allowing the CUT to be detected.
• Critical Limitation: Unacceptable false alarm rate at clutter edges, as the smaller window may severely underestimate the true noise floor in the high-clutter region.

A rank-based approach that replaces arithmetic averaging with a k-th order statistic of the sorted reference window samples.

• Mechanism: Sorts the amplitudes of all reference cells and selects the k-th largest value as the noise estimate, multiplied by a scaling factor.
• Robustness: Inherently immune to up to (N - k) interfering targets within the reference window, where N is the total number of reference cells.
• Typical Configuration: Often uses the 75th percentile (e.g., k = 3N/4) to balance multi-target tolerance with estimator variance. Higher computational cost than CA-CFAR due to the sorting operation.

An adaptive variant that explicitly identifies and excises interfering targets from the reference window before computing the noise estimate.

• Mechanism: Applies a pre-threshold to censor (remove) reference cells whose power exceeds a cutoff, then averages only the remaining cells assumed to contain pure noise.
• Iterative Censoring: Advanced implementations iteratively refine the censoring threshold based on the current noise estimate to prevent self-masking.
• Application: Highly effective in dense signal environments where multiple non-homogeneities must be rejected without prior knowledge of their number or positions.

An adaptive composite algorithm that dynamically selects between CA-CFAR, GO-CFAR, and SO-CFAR based on real-time statistical tests of the reference window.

• Mechanism: Computes a variability index (ratio of variance to mean) and a mean ratio between the leading and lagging windows to classify the environment as homogeneous, clutter edge, or multiple-target.
• Decision Logic: A lookup table maps the two test outcomes to the optimal CFAR variant for the current local conditions.
• Advantage: Provides robust detection across heterogeneous environments without requiring a priori knowledge of the interference structure.

The most common CFAR variant that estimates the local noise floor by averaging the power of adjacent range or frequency bins surrounding the cell under test.

• Mechanism: A sliding window of reference cells is divided into leading and lagging groups, with guard cells placed immediately adjacent to the test cell to prevent signal leakage.
• Key assumption: The background interference is homogeneous and identically distributed across the reference window.
• Weakness: Suffers from masking effects in non-homogeneous clutter, where a strong interfering target in the reference window raises the threshold and obscures a weaker nearby signal.

A robust CFAR technique that sorts the reference cell values and selects the k-th ordered sample to set the threshold, rather than using the arithmetic mean.

• Advantage: Highly resilient to multiple interfering targets in the reference window, as outlier cells are naturally pushed to higher ranks and ignored.
• Trade-off: Incurs a small CFAR loss in homogeneous noise compared to CA-CFAR due to the inefficiency of order statistics versus averaging.
• Typical rank: The multiplier is often set to select the 75th percentile of the sorted reference window to balance detection sensitivity and false alarm control.

Two adaptive variants designed for specific non-homogeneous environments by comparing the noise estimates from the leading and lagging reference sub-windows independently.

• GO-CFAR (Greatest-Of): Selects the maximum of the two sub-window averages. Designed to suppress false alarms at clutter edges, where one side of the window contains high-power returns and the other contains only noise.
• SO-CFAR (Smallest-Of): Selects the minimum of the two sub-window averages. Optimized for detecting a weak target closely spaced to a stronger target, where one sub-window is contaminated by the strong return.

The adaptive power level computed by multiplying the estimated noise floor by a scaling factor derived from the desired probability of false alarm (Pfa).

• Formula: T = α × μ̂, where μ̂ is the local noise estimate and α is the threshold multiplier determined analytically from the assumed noise distribution and target Pfa.
• Rayleigh envelope: For square-law detection in complex Gaussian noise, α = N × (Pfa^(-1/N) - 1), where N is the number of reference cells.
• Practical Pfa values: Radar systems typically operate at Pfa = 10⁻⁴ to 10⁻⁶, while spectrum sensing applications may tolerate 10⁻² to 10⁻³.

Buffer cells placed immediately adjacent to the cell under test (CUT) that are excluded from the noise power estimation to prevent signal energy from contaminating the threshold calculation.

• Purpose: A modulated signal's energy often spills over into neighboring range or frequency bins due to pulse shaping, Doppler processing sidelobes, or spectral leakage.
• Configuration: The number of guard cells is matched to the expected bandwidth or duration of the target signal. Too few guard cells cause self-masking; too many reduce the effective reference window size and increase CFAR loss.
• Typical values: 1-4 guard cells on each side for radar range processing; wider guard bands for frequency-domain CFAR in spectrum sensing.

The additional signal-to-noise ratio (SNR) required by a CFAR detector to achieve the same detection probability as an ideal fixed-threshold detector with perfect noise knowledge.

• Cause: Finite reference window size introduces estimation variance in the noise floor, requiring a higher threshold multiplier to maintain the target Pfa, which in turn demands more signal power.
• CA-CFAR loss: Approximately 1-3 dB for typical reference window sizes (16-32 cells), decreasing as the number of reference cells increases.
• OS-CFAR penalty: An additional 0.3-0.5 dB loss compared to CA-CFAR in homogeneous noise, accepted as the cost of robustness to interfering targets.