Constant False Alarm Rate (CFAR)

The foundational approach that computes the noise threshold by averaging the power of adjacent reference cells surrounding the Cell Under Test (CUT). It is the optimal detector in a homogeneous, exponentially distributed background.

• Mechanism: Sums the outputs of N reference cells and multiplies by a scaling factor derived from the desired false alarm rate.
• Key Weakness: Suffers from masking when multiple targets are in close proximity, as the strong returns from adjacent targets inflate the noise estimate.
• Application: Ideal for clear environments with isolated targets and uniform clutter.

Designed to mitigate false alarms at clutter edges, this variant calculates the noise estimate independently for the leading and lagging reference windows and selects the larger of the two.

• Clutter Edge Robustness: Prevents the threshold from dropping in the clear region adjacent to a high-power clutter patch, which would otherwise cause excessive false alarms.
• Trade-off: Exhibits degraded detection performance for closely spaced targets compared to CA-CFAR.
• Use Case: Environments with distinct transitions between clutter densities, such as land-sea interfaces.

An inverse logic to GO-CFAR that selects the smaller of the two independent window averages. This technique is specifically engineered to resolve multiple targets separated by less than the reference window length.

• Anti-Masking: Prevents a strong target in one window from suppressing the detection of a weaker target in the adjacent CUT.
• Trade-off: Highly susceptible to false alarms at clutter edges because it selects the lower noise estimate.
• Application: Dense target environments where resolving individual objects is critical.

A rank-based method that sorts the reference cell values in ascending order and selects the k-th ordered sample as the noise estimate. This discards outlier interferers that corrupt mean-based estimators.

• Mechanism: The threshold multiplier is derived from the order statistic's probability density function.
• Robustness: Inherently immune to up to (N - k) interfering targets within the reference window.
• Implementation: Requires a sorting network, increasing computational complexity on FPGAs but providing superior performance in multi-target scenarios.

A two-pass algorithm that first identifies and excises (censors) outlier cells exceeding a pre-threshold before computing the mean of the remaining samples. This dynamically cleans the reference window.

• Adaptive Censoring: Iteratively removes interfering targets to prevent the noise floor estimate from being biased upward.
• Censoring Logic: Often uses a fixed number of highest samples to remove or a data-dependent hypothesis test.
• Advantage: Maintains near-optimal CA-CFAR performance in homogeneous environments while automatically adapting to non-homogeneity.

An intelligent composite algorithm that dynamically switches between CA-CFAR, GO-CFAR, and SO-CFAR based on a statistical variability index and a mean ratio test computed on the reference windows.

• Decision Logic: The variability index detects non-homogeneity, while the mean ratio detects clutter edges.
• Adaptive Selection: Automatically selects SO-CFAR for multiple targets, GO-CFAR for clutter edges, and CA-CFAR for homogeneous noise.
• Complexity: Requires real-time statistical hypothesis testing but provides robust detection without a priori environmental knowledge.

The foundational process that CFAR relies upon. Noise floor estimation determines the background power level of a receiver in the absence of a signal. CFAR algorithms continuously update this estimate to adapt the detection threshold. Common methods include:

• Cell-averaging: Averaging the power in adjacent range/Doppler bins
• Ordered statistics: Selecting the k-th smallest value from reference cells
• Greatest-of / Smallest-of: Choosing the max or min of leading and lagging windows Accurate estimation is critical; an underestimated noise floor causes false alarms, while an overestimated floor masks real targets.

The most common CFAR variant. CA-CFAR computes the detection threshold by averaging the power in a set of reference cells surrounding the cell under test (CUT). A guard band of empty cells immediately adjacent to the CUT prevents target energy from leaking into the noise estimate. The threshold is the average multiplied by a scaling factor derived from the desired probability of false alarm. CA-CFAR is optimal in homogeneous noise but degrades in clutter edges or multi-target environments.

A robust alternative to CA-CFAR designed for multi-target environments. Instead of averaging, OS-CFAR sorts the reference cell values and selects the k-th ordered sample as the noise estimate. This makes it immune to interference from closely spaced targets that would inflate a cell-averaging estimate. The rank parameter k is typically chosen as 3/4 of the total reference window size, providing strong false alarm control even when up to N - k interfering targets are present in the reference window.

The key performance metric that CFAR is designed to hold constant. PFA is the probability that noise alone exceeds the detection threshold, generating a spurious target declaration. In radar systems, typical PFA values range from 10⁻⁴ to 10⁻⁶. The relationship between PFA, the threshold multiplier α, and the number of reference cells N is given by: α = N × (PFA^(-1/N) - 1). Maintaining a constant PFA ensures predictable system behavior regardless of changing environmental noise.

Critical buffer cells placed immediately adjacent to the cell under test (CUT) that are excluded from the noise estimation window. Guard bands prevent target self-interference—where a strong target's sidelobe energy spills into reference cells and artificially raises the noise estimate, suppressing detection. The width of the guard band is matched to the expected target spread in the given domain:

• Range domain: Matched to the compressed pulse width
• Doppler domain: Matched to the target's spectral broadening Insufficient guard bands cause self-masking of extended targets.

A known failure mode of basic CFAR algorithms. At a clutter edge—a sharp transition from low to high background power—standard CA-CFAR exhibits degraded performance:

• Leading edge: False alarms spike as the window enters high clutter
• Trailing edge: Detection probability drops as the window still contains high clutter GO-CFAR (Greatest-Of) mitigates false alarms at leading edges by selecting the larger of the leading and lagging window averages. SO-CFAR (Smallest-Of) improves trailing-edge detection by selecting the smaller average, at the cost of higher false alarm rate.