Motion Compensation for Multiple-Input-Multiple-Output Inverse Synthetic Aperture Imaging of Automotive Targets

Inverse synthetic aperture radar (ISAR) images generated from single-channel automotive radar data provide critical information about the shape and size of automotive targets. However, the quality of ISAR images degrades due to road clutter and when translational and higher order rotational motions of the targets are not suitably compensated. One method to enhance the signal-to-clutter-and-noise ratio (SCNR) of the systems is to leverage the advantages of the multiple-input-multiple-output (MIMO) framework available in commercial automotive radars to generate MIMO-ISAR images. While substantial research has been devoted to motion compensation of single-channel ISAR images, the effectiveness of these methods for MIMO-ISAR has not been studied extensively. This paper analyzes the performance of three popular motion compensation techniques - entropy minimization, cross-correlation, and phase gradient autofocus - on MIMO-ISAR. The algorithms are evaluated on the measurement data collected using Texas Instruments millimeter-wave MIMO radar. The results indicate that the cross-correlation MOCOMP performs better than the other two MOCOMP algorithms in the MIMO configuration, with an overall improvement of 36%.

💡 Research Summary

This paper investigates motion compensation (MOCOMP) techniques for multiple‑input‑multiple‑output (MIMO) inverse synthetic aperture radar (ISAR) imaging of automotive targets. While single‑channel (SISO) ISAR has been extensively studied, the performance of popular fine‑level MOCOMP algorithms in a MIMO configuration has not been thoroughly examined. The authors evaluate three widely used algorithms—entropy minimization (EM), cross‑correlation (CCR), and phase‑gradient autofocus (PGA)—using real measurement data collected with a Texas Instruments WR1843 millimeter‑wave MIMO radar operating at 77 GHz with a 2 GHz bandwidth.

The radar is set up in a monostatic TDM‑MIMO mode with 12 virtual channels (P × Q). FMCW chirps are transmitted sequentially from each of the P transmit elements, and the received signals from all Q receive elements are processed jointly. After stretch processing and a 2‑D FFT, range‑Doppler matrices are generated for each TX‑RX pair, then non‑coherently summed to form a MIMO‑ISAR image for each coherent processing interval (CPI).

Motion compensation is performed in two stages. Coarse MOCOMP removes the dominant range‑walk by phase‑correcting the strongest scatterer across slow‑time, effectively fixing the target’s range position. Fine MOCOMP then estimates translational velocity, acceleration, and higher‑order rotational parameters (yaw, pitch, roll) and applies phase corrections. The three fine‑level algorithms are implemented as follows:

1. Entropy Minimization (EM) iteratively searches for motion parameters that minimize the entropy of the ISAR image, aiming to concentrate energy into a compact region. In practice, EM suffers from slow convergence and produces noticeable smearing when applied to noisy real‑world data.

2. Cross‑Correlation (CCR) computes the correlation between successive echoes (or across channels) to estimate phase misalignment. The authors run 97 iterations, observing that CCR rapidly reduces the noise floor and sharpens target edges in many frames, especially when the multiple virtual channels provide redundant information. However, convergence is not uniform; a few frames exhibit residual errors.

3. Phase‑Gradient Autofocus (PGA) directly estimates the phase gradient between consecutive chirps and compensates it. PGA converges faster than EM and yields clearer target signatures, but it leaves prominent side‑lobes along the Doppler axis, limiting overall image contrast.

The experimental scenario involves a Hyundai Santro (3.6 × 1.6 × 1.6 m) performing a U‑turn trajectory while the radar remains stationary at the origin. The measurement set includes 128 slow‑time samples and 256 fast‑time samples per CPI, with a maximum unambiguous range of 34.4 m and a velocity limit of 5 m/s. Ground truth is recorded with an auxiliary camera.

Results are presented for both SISO and MIMO configurations across several time instants (8.1 s to 9.0 s). In the SISO case, all three algorithms improve signal‑to‑clutter‑and‑noise ratio (SCNR) modestly, but EM shows severe smearing, PGA exhibits strong side‑lobes, and CCR provides only a modest 6 % SCNR gain. In the MIMO case, the virtual array’s spatial diversity amplifies the benefits of fine MOCOMP. The coefficient of variation (CV) of the noise floor—used as a quantitative metric—demonstrates that CCR achieves a 36 % improvement in SCNR relative to the un‑compensated MIMO image, outperforming EM (0.1 % improvement) and PGA (3.57 % improvement). Visual inspection of the range‑Doppler plots confirms that CCR yields the most focused target energy and the lowest clutter level for most frames, though occasional frames still suffer from residual motion artifacts.

The study also notes the appearance of “ghost” targets caused by multipath reflections, indicating that future work should incorporate multipath mitigation strategies. The authors conclude that motion compensation algorithms validated on SISO ISAR cannot be directly transferred to MIMO ISAR without re‑evaluation. Cross‑correlation based compensation leverages the redundancy across virtual channels and is the most effective technique for enhancing MIMO‑ISAR image quality under realistic automotive scenarios.

Future research directions include reducing computational complexity for real‑time implementation, integrating higher‑order rotational modeling (yaw, pitch, roll) into a unified compensation framework, and extending validation to diverse vehicle types, speeds, and clutter environments. The paper provides a solid experimental foundation for advancing high‑resolution automotive radar imaging using MIMO configurations and sophisticated motion compensation.