CA-CFAR Detector

• The paper presents the CA-CFAR detector as a method that adapts detection thresholds based on local noise estimates using guard and reference cells, ensuring a constant false alarm rate.
• It introduces a continuous relaxation via a sigmoid function, enabling the CA-CFAR mechanism to be integrated into end-to-end differentiable neural network pipelines for optimized detection metrics.
• Empirical evaluations, particularly on automotive radar data, show that using separable convolutions and balanced cross-entropy loss significantly improves F1-scores compared to classical detection methods.

The cell-averaging constant false alarm rate (CA-CFAR) detector is a canonical approach for peak detection in radar signal processing, providing adaptive thresholding based on local noise estimates to maintain a specified probability of false alarm (Pfa). In object detection contexts, notably automotive radar, the CA-CFAR mechanism forms the algorithmic backbone for differentiating target returns from noise across multidimensional range–Doppler–angle (RDA) maps, and its recent integration into differentiable neural-network pipelines has allowed for direct optimization of detection metrics within learning systems (Oswald et al., 2023).

1. Classical CA-CFAR Formulation

The CA-CFAR detector operates on complex-valued radar returns $x(r)$ within range cells indexed by $r$ across an RDA map. The algorithm designates each cell as a cell-under-test (CUT) and surrounds it with $G$ guard cells—omitted from noise estimation to avoid contamination by the target—and $N$ reference cells on each side for local noise power estimation. The reference cell set is defined by

$$\Omega_{\mathrm{ref}} = \{ i \; | \; G < |i| \leq G + N \}.$$

The local noise power at cell $r$ is estimated as

$$\widehat{P}_n(r) = \frac{1}{2N}\sum_{i \in \Omega_{\mathrm{ref}}} |x(r+i)|^2,$$

and the detection threshold is

$T(r) = \alpha \widehat{P}_n(r),$

where $\alpha$ is set according to desired Pfa. The binary decision at each cell is conventionally

$$y(r) = \begin{cases} 1, & |x(r)|^2 > T(r), \ 0, & \text{otherwise}. \end{cases}$$

2. Continuous Relaxation for Differentiable Learning

To embed CA-CFAR within end-to-end neural networks, the rigid thresholding step is relaxed via a sigmoid function, allowing backpropagation of gradients. The local signal-to-interference-plus-noise ratio (SINR) is

$$\mathrm{SINR}(r) = \frac{|x(r)|^2}{\widehat{P}_n(r)}$$

with a threshold $\beta = \alpha$. The relaxed output is

$$C_{\mathrm{CFAR}}(r) = \sigma\left(\frac{\mathrm{SINR}(r)-\beta}{\tau}\right) = \frac{1}{1+\exp\left(\frac{\beta-\mathrm{SINR}(r)}{\tau}\right)},$$

where $\tau > 0$ modulates the smoothness of the transition. In terms of the raw signal and noise estimate,

$$C_{\mathrm{CFAR}}(r) = \sigma\left(\frac{|x(r)|^2-\beta\,\widehat{P}_n(r)}{\tau\,\widehat{P}_n(r)}\right).$$

All arithmetic operations—squaring, summation, division—are differentiable, ensuring that gradients propagate through the CA-CFAR block in neural architectures.

3. Loss Function Integration and Training

Within neural-network-based detection tasks, both the predicted RDA map $\hat{S}$ and the ground-truth clean map $S$ are processed by the relaxed CA-CFAR block:

$$\hat{y}(r) = C_{\mathrm{CFAR}}(\hat{S}(r)), \quad y(r) = C_{\mathrm{CFAR}}(S(r)), \; (y, \hat{y} \in [0,1])$$

Comparisons employ Balanced Cross-Entropy (BCE) to mitigate class imbalance:

$$\mathrm{BCE}(y, \hat{y}; \alpha_w) = -\,\alpha_w\,y\log\hat{y} - (1-\alpha_w)(1-y)\log(1-\hat{y})$$

with $\alpha_w \in (0,1)$, empirically optimized to $0.75$. The global loss aggregates the BCE over all RDA cells; standard batch-normalization weight decay suffices for regularization.

4. Neural-Network Architectures and Kernel Efficiency

The architecture, termed AENN, accepts a $96\times96\times16$ complex tensor corresponding to range, Doppler, and angle. All convolutions are complex-valued, stride-one, with zero-padding. The layer pipeline comprises:

• Layer 1: 4 output channels, kernel $[K_A \times K_D \times K_R]$, complex ReLU, complex BatchNorm
• Layer 2: 2 output channels, same kernel, ReLU, BatchNorm
• Layer 3: 1 output channel, same kernel, no nonlinearity

Kernel representation significantly affects parameter count:

• Generic 3D kernel ($K^3$ complex parameters/filter)
• Separable kernel (factorized into 3 axes, $3K$ parameters/filter)

For $K=3$: | Kernel Type | Parameter Count (real-valued) | Multiplies per location | |----------------|-------------------------------|------------------------| | Generic 3D | 800 | $K^3$ | | Separable | 296 | $3K$ |

The separable formulation achieves marked reduction in parameterization and computational load, with quantifiable impact on storage and speed.

5. Empirical Performance and Ablation Results

On real-world automotive radar data, quantifying detection using F1-score (tolerance $\pm3$ range, $\pm3$ Doppler, $\pm1$ angle):

• AENN + BCE + separable $[7\times7\times7]$, 800 params: F1 = 0.921
• AENN (BCE) generic $[3\times3\times3]$, 800 params: F1 = 0.844
• Classical methods: zeroing F1 = 0.655, ramp filtering F1 = 0.551, IMAT F1 = 0.513

No explicit ROC curves are plotted, but the F1-score indicates superior balance of true positives and false alarms in neural approaches. Training with simple magnitude-MSE or full MSE yields F1 $<$ 0.85, establishing the advantage of CA-CFAR relaxation with BCE for detection objectives.

6. Algorithmic Integration and Significance

Embedding a differentiable CA-CFAR detector within end-to-end learning loops allows neural networks to directly optimize object detection metrics under a fixed CFAR regime. This approach decouples performance from pure signal regression and aligns detection with statistical false-alarm constraints. The use of separable convolutions supports parameter and computational efficiency without sacrificing accuracy; in practice, detection accuracy was preserved or improved despite model downsizing to several hundred parameters. This paradigm is validated through rigorous experimentation and benchmarks against classical interference mitigation strategies (Oswald et al., 2023).

7. Context within Radar Object Detection Methodologies

The CA-CFAR detector remains fundamental in radar object detection, balancing adaptivity and analytical tractability. Its continuous relaxation and integration into modern deep learning illustrate a convergence between statistical signal processing and neural-network-driven methods. This suggests a generalizable template for embedding legacy detection algorithms as differentiable components within novel architectures, facilitating performance gains while preserving interpretability and principled thresholding.