Edge-Based, MmWave Motion Compensation

Note: Please view the desktop site in order to see images.

In recent years, the technological community has witnessed remarkable simultaneous innovation in the areas of 1) robust and dynamically configurable silicon-based CMOS mmWave radar sensors, 2) powerful processing engines operating at the edge, and 3) increasingly generalizable machine learning models. Consolidated, the advances in these three domains have created an eminent opportunity for the creation of truly autonomous mmWave edge perception devices to an extent heretofore unexplored. While edge sensing is by no means a novel concept in a modern age of distributed IoT networks, the inherent tradeoff between computational capacity and power consumption has severely limited the capabilities of battery-powered sensors at the edge. Moreover, while cloud computing and multi-core processors offer an undeniable advantage in terms of sheer computation efficiency and scalability, the reality is that many detection scenarios require real-time, high-resolution perception right at the edge, without the latency incurred by bulk post-processing the data collected by passive mmWave radar sensors.

Unfortunately, the perception and resolution capacity of edge devices is inevitably constrained by data degradation in the presence of parasitic platform motion, an impairment exacerbated for highly sensitive mmWave devices as a result of their phase-based coherent processing. Edge sensors mounted upon automotive vehicles, drones, and industrial machinery, for instance, are subject to engine or motor vibration in addition to any other nonharmonic transient movement aberrations produced by bearing faults or external forces (e.g., road or air turbulence). Similarly, wearable devices for motion capture, mobility aid, and health monitoring applications (e.g., blood pressure monitoring, vital signs sensing, etc.) are susceptible to the effects of user motion, both deterministic (e.g., harmonic), and nondeterministic. To address and correct for such parasitic motion, we have developed several real-time-compatible, low-complexity signal-processing algorithms, demonstrated and verified on measured mmWave radar data.

Fusion-Based Vibration Compensation

In [1], we have demonstrated a real-time solution for single-mode vibratory motion correction, proposing a novel frequency-domain deconvolution-based signal processing algorithm operating at the FMCW radar frame level. Fusing synchronized data from an auxiliary IMU, we have demonstrated up to 40 dB motion suppression capacity on a < 100-ms correction time scale. This short-time-scale, latency-free algorithm may be easily implemented in real time on a computation-constrained mmWave edge device, significantly improving sensing resolution and accuracy through autonomous, in-sensor/on-chip motion compensation.

Anchor-Based Motion Compensation

In [2], we have extended our work to more generalized, complex platform motion comprising both multimode harmonic motion and nondeterministic transient spur components.. In particular, we have developed a unique radar anchor point data fusion algorithm capable of correcting complex platform motion, either complementing or eliminating the need for an auxiliary sensor (e.g., IMU) and thus increasing system redundancy and reliability at a reduced hardware overhead.  Measurements and evaluation using the designed motion generation and radar data collection hardware setup demonstrate comparable motion suppression performance to that of our previous work, despite the increased motion complexity and absence of auxiliary sensor data.

Publications

[1] N. Poole, S. Hor, and A. Arbabian, “A Real-Time, Frame-Level Platform Vibration Compensation Approach for mmWave Radar Systems,” 18th European Radar Conference (EuRAD), 2021.

[2] N. Poole and A. Arbabian, “Anchor-Based, Real-Time Motion Compensation for High-Resolution mmWave Radar,” IEEE Journal of Microwaves, vol. 4, no. 3, pp. 440-458, Jul. 2024.