Monostatic Automotive SAR Systems

• Monostatic automotive systems are radar architectures featuring co-located TX and RX elements on vehicles, leveraging SAR for detailed 3D mapping.
• They utilize FMCW MIMO arrays at 77 GHz with time-division multiplexing and ego-motion compensation to achieve sub-degree azimuth and centimeter-level elevation accuracy.
• These systems balance advanced signal processing and computational constraints to effectively detect static objects and enhance scene perception in autonomous driving.

Monostatic automotive systems refer to radar architectures where transmitting and receiving antennas are co-located on a moving vehicle, utilizing synthetic aperture radar (SAR) and related interferometric extensions to achieve enhanced angular resolution and three-dimensional (3D) mapping for automated driving applications. These systems are characterized by integration of compact millimeter-wave (mmWave) MIMO arrays on vehicle platforms, advanced signal processing pipelines, and the selective leveraging of ego-motion to synthetically extend aperture length for improved performance in static object detection, mapping, and scene perception.

1. System Architectures and Signal Models

Monostatic automotive SAR platforms typically employ frequency-modulated continuous-wave (FMCW) MIMO radars at 77 GHz with multiple transmit (TX) and receive (RX) elements. The physical array is compact (e.g., ∼10 cm aperture), often arranged in two elevation layers with vertical baseline spacing of $D_v = \lambda/4$ (where $\lambda$ is the radar wavelength). Multiple TX/RX arrangements yield virtual elements through time-division multiplexing (TDM), commonly resulting in 12 virtual monostatic channels (“VXs”) (Kabuli et al., 14 Jan 2025).

During operation, a moving vehicle forms a synthetic aperture by traversing a segment of trajectory, accumulating radar returns over $N$ frames (each of duration $T_f$), thus achieving a synthetic aperture length of $\sum_n v_n T_f$ for ego-velocity $v_n$ per frame (Bialer et al., 2022). The received signals are mixed with the transmitted chirp and digitized, yielding a data cube $y_n(r, \phi, f_D)$ structured along range, angle, and Doppler bins—this is foundational for subsequent SAR focusing.

2. Synthetic Aperture and SAR Focusing

Synthetic aperture formation in monostatic automotive SAR exploits platform motion to improve angular resolution far beyond that achievable by the physical array. For static targets, range migration is compensated by coherently summing matched-filtered outputs along hypothesized trajectories corresponding to specific range and bearing parameters. The SAR output for hypothesized reflector position $(\tilde\gamma, \tilde\theta)$ is computed as

$$\mu(\tilde\gamma,\tilde\theta,\tilde V) \propto \left| \sum_{n=0}^{N-1} y_n(\tilde\gamma_n,\tilde\theta,\tilde v_n) \, e^{-j\frac{4\pi}{\lambda} r_n(\tilde V,\tilde\theta)} \right|,$$

with $$\tilde\gamma_n = \tilde\gamma - r_n(\tilde V, \tilde\theta)$$ and $$r_n(\tilde V, \tilde\theta) = T_f \sum_{k=0}^n p^T(\tilde\theta)\tilde v_k$$ where $p(\theta) = [\sin\theta,\,\cos\theta]^T$ (Bialer et al., 2022). Fast Back-Projection (FBP) is the standard focusing algorithm, integrating echo contributions for each candidate scene point based on instantaneous geometry and ego-localization (Kabuli et al., 14 Jan 2025).

Range resolution is governed by the bandwidth $B$ as $\Delta R = c/(2B)$, and azimuth resolution, or SAR angular resolution, scales as $$\Psi_{\mathrm{SAR}}(\theta) = \lambda/(L\sin\theta)$$, with $L$ the synthetic aperture length. For $L=1\,\text{m}$ and broadside ($\sin\theta\approx1$), resolutions below $0.25^\circ$ are feasible (Kabuli et al., 14 Jan 2025).

3. Ego-Velocity Estimation and Angle Error Analysis

SAR imaging performance relies critically on precise vehicle velocity estimation at each frame, as velocity errors directly induce angle estimation errors and SAR defocusing. Radar-only velocity estimation is implemented by leveraging static-object Doppler returns, formulating a frame-wise least-squares problem: $$Q(\tilde v_n) = \frac{1}{2} \| G(\phi_n)\tilde v_n - f_n \|_2^2,$$ where $f_n = [f_{n,1},...,f_{n,K}]^T$ (Dopplers), $G(\phi_n)$ constructed from target angles, and $K$ denotes the number of static points-of-interest (PPIs). The closed-form minimizer is

$$\hat v_n = \left[G(\phi_n)^T G(\phi_n)\right]^{-1} G(\phi_n)^T f_n$$

(Bialer et al., 2022).

For SAR systems using radar-estimated velocity, angle error variance is analytically characterized. In straight-motion, large-$N$, large-$K$ regimes, the SAR angle error variance simplifies to

$$\mathrm{Var}(\delta\theta) \approx \frac{\sigma_f^2\lambda^2+\sigma_\phi^2 v_y^2(1+2\sin^2\theta)} {2K\omega(N)v_y^2\sin^2\theta}$$

where $\sigma_\phi^2$ is array angle-estimate variance, $\sigma_f^2$ Doppler variance, $\omega(N)$ grows with number of frames $N$, and $v_y$ is forward vehicle speed. Notably, variance diverges as $\sin\theta\to 0$ (boresight) and decays with increasing $K$ and $N$ (Bialer et al., 2022).

4. Three-Dimensional Scene Mapping with Interferometric SAR

Elevation resolution is not feasible using single-layer monostatic arrays alone. To address this, interferometric SAR (InSAR) extends monostatic SAR by combining pairs of vertically separated virtual elements, extracting height from inter-channel phase. After FBP focusing per VX, the interferometric phase $\Delta\psi(u,v)$ is computed for each vertical baseline: $$\Delta\psi(u,v) = \angle \left\{ S_0(u,v) S_1^*(u,v) \right\}$$ where $S_i(u,v)$ is the complex SAR pixel for VX$_i$ (Kabuli et al., 14 Jan 2025).

The elevation angle $\phi$ is derived by

$$\phi = \sin^{-1}\left( \frac{\lambda}{4\pi D_v} \Delta\psi \right),\quad D_v\le\frac{\lambda}{4}$$

Height estimation follows through spherical-to-Cartesian conversion using the triangulated range and azimuth. Point clouds are generated by thresholding on signal-to-noise ratio (SNR), filtering on phase variance, and merging left/right radar outputs after geo-registration using GNSS/INS pose data (Kabuli et al., 14 Jan 2025).

In controlled conditions, InSAR achieves sub-cm elevation accuracy (absolute errors ≤9 mm on 33/63 cm reflectors; ≈14 mm at ground), with robust discrimination of typical automotive scene features (e.g., cars, curbs, buildings, vegetation) in field settings (Kabuli et al., 14 Jan 2025).

5. Performance Regimes, Trade-offs, and Limitations

Key performance parameters and their effects on SAR angular error variance are summarized below:

Parameter	Performance Effect	Direction
$\sigma_\phi$ (angle var)	$\uparrow$ ⇒ $\mathrm{Var}(\delta\theta)\uparrow$	Detrimental
$\sigma_f$ (Doppler var)	$\uparrow$ ⇒ $\mathrm{Var}(\delta\theta)\uparrow$	Detrimental
$K$ (static PPIs)	$\uparrow$ ⇒ $\mathrm{Var}(\delta\theta)\downarrow$	Beneficial
$N$ (integration frames)	$\uparrow$ ⇒ $\mathrm{Var}(\delta\theta)\downarrow$	Beneficial (latency cost)
$v_y$ (speed)	$\uparrow$ ⇒ focus $\uparrow$ (to plateau)	Mixed
Target angle $\theta$	$\sin\theta\downarrow$ ⇒ $\mathrm{Var}(\delta\theta)\uparrow$	Detrimental at boresight

Simulation results show that monostatic automotive SAR (with radar-only ego-velocity estimation) yields significant angular resolution gain (e.g., 1° down to 0.2–0.4°) only under conditions of abundant static reflectors ($K\gtrsim 5$–10), adequate speed ($v_y\gtrsim 10$ m/s), sufficient integration ($N\gtrsim 5$), and target angles offset from boresight ($|\theta|\gtrsim 10^\circ$). Maximum resolution gain (≈3× improvement) is observed at $|\theta| > 40^\circ$, $K\geq 15$, and $N\geq 10$ (Bialer et al., 2022).

A critical limitation is that for low $v_y$, sparse static scenes (low $K$), or short apertures (low $N$), SAR can perform worse than the short-aperture array baseline. Only static-object SAR is feasible—moving targets are defocused. Elevation extraction is fundamentally limited by the vertical baseline $D_v$; finer height granularity or broader ambiguous height range requires a larger or multi-orientation vertical aperture (Kabuli et al., 14 Jan 2025).

6. Computational Complexity and Implementation Constraints

SAR back-projection and range migration algorithms entail substantial computation, especially over fine grids in $(r,\theta)$ for $N$ frames, with additional phase-compensation for motion. Interferometric phase calculation is linear in the number of pixels. For example, a 30 m × 30 m SAR image at 4 cm pixel spacing can be processed offline for a 12 VX system in under 1 s on a mid-range CPU; real-time implementation is straightforward on GPUs or FPGAs (Kabuli et al., 14 Jan 2025).

Velocity fitting per frame requires solving a $2\times2$ least squares system per $T_f$. The principal practical trade-offs involve latency (detection delayed by $N\times T_f$) and compute load, motivating selective deployment in static-rich, moderate-speed scenarios such as urban parking lots. Fusion with GNSS/INS may reduce velocity estimation errors but at higher hardware cost and complexity (Bialer et al., 2022).

7. Comparative Analysis and Applications

Compared to conventional monostatic arrays, which offer 5–10° azimuth resolution and limited or no elevation accuracy, monostatic automotive SAR achieves 0.18 m range and $<0.25^\circ$ azimuth resolution, while SAR + InSAR adds sub-degree elevation sensitivity (Δφ ≈ 0.1° at high SNR), supporting centimeter-level 3D mapping (Kabuli et al., 14 Jan 2025). Validation in agricultural and urban scenarios demonstrates robust mapping of objects (cars, trees, buildings, pedestrians) up to 30 m range with high elevation and cross-range fidelity, with practical avoidance of artifacts through phase-masking and geometric correction.

These systems enable high-fidelity, low-latency 3D point cloud generation for perception tasks in autonomous vehicles, under constraints of moderate computation and sensor cost. Nonetheless, their efficacy is scenario-dependent, and use must be tailored to situations where static scene structure, platform speed, and infrastructure justify the operational complexity (Bialer et al., 2022, Kabuli et al., 14 Jan 2025).