Constant False Alarm Rate (CFAR) Detector

CFAR detectors dynamically calculate a detection threshold by estimating the local noise power from surrounding reference cells. Unlike a fixed threshold, which fails in fluctuating noise environments, the adaptive threshold rises and falls with the interference floor. The algorithm multiplies the estimated noise level by a scaling factor derived from the desired false alarm probability, ensuring that a target signal must exceed the local background by a statistically significant margin to trigger a detection.

The most common variant, Cell-Averaging CFAR, computes the noise estimate by averaging the power in a set of leading and lagging reference windows surrounding the Cell Under Test (CUT). Guard cells immediately adjacent to the CUT are excluded to prevent target signal energy from leaking into the noise estimate and biasing the threshold upward. This approach is optimal in homogeneous noise but suffers from masking effects when multiple targets are present in the reference window.

Ordered-Statistic CFAR improves performance in multi-target and clutter-edge environments by sorting the reference cell values and selecting the k-th ordered sample as the noise estimate. This technique inherently rejects outlier interferers that would inflate a simple average. By choosing an appropriate rank, typically around the 75th percentile, OS-CFAR maintains robust detection of weaker targets near stronger returns without the catastrophic threshold elevation that plagues CA-CFAR in non-homogeneous scenes.

Guard cells are the buffer samples immediately adjacent to the Cell Under Test that are excluded from the noise power calculation. Their purpose is to prevent target spillover—where a distributed or moving target's energy contaminates the reference window—from raising the threshold and suppressing detection. The number of guard cells must be tuned to the expected target bandwidth and range migration. Too few guard cells cause self-masking; too many reduce the number of effective reference samples and degrade the noise estimate's statistical accuracy.

CFAR loss quantifies the additional Signal-to-Noise Ratio (SNR) required by an adaptive detector to achieve the same detection probability as an ideal fixed-threshold detector with perfect noise knowledge. This loss arises from the finite number of reference cells used to estimate the noise, introducing estimation variance. Typical CFAR loss ranges from 1 to 3 dB depending on the window size and algorithm variant. Larger reference windows reduce loss but increase computational load and degrade performance in non-stationary interference.

In modern cognitive radio and spectrum awareness pipelines, CFAR detectors serve as a critical front-end to deep learning classifiers. By isolating potential signals from raw noise, CFAR reduces the data volume fed into computationally expensive neural networks and provides region-of-interest proposals. The detected signal segments—extracted as IQ samples or spectrogram crops—are then passed to downstream models for modulation recognition, RF fingerprinting, or interference classification, forming a two-stage sensing architecture.