Real-Time, High-Resolution mmWave Radar Sensing

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

Across both industry and academia, IoT sensor networks have benefited from the development of high-resolution silicon-based sensor devices and post-processing engines for accurate information extraction and perception. Existing implementations, however, entail substantial resource/power consumption and processing latency, rendering them infeasible for edge-based sensors operating without any back-end cloud computing or post-processing. Our lab addresses this shortcoming for edge devices by exploring the design of analog/mixed-signal and RF integrated circuits for real-time, high-resolution mmWave radar sensing and processing applications. In particular, our research explores the use of mmWave radar sensor fusion and closed-loop adaptive sensing circuit architectures, with custom circuits hardware and back-end ML/signal processing algorithms designed in parallel to enable energy-efficient, high-resolution real-time radar sensing.

Sensor Fusion Chip for High-Resolution mmWave Motion Compensation

One of the most prevalent issues limiting high-resolution mmWave perception is the parasitic motion of the radar platform itself. When mounted upon a non-stationary platform such as a vibrating automotive engine, airborne drone, or wearable device, the mmWave radar sensor generates corrupted data which hampers accurate motion parameter estimation. To combat this issue, we have designed a 130nm custom analog/mixed-signal IC implementing inertial sensor-fusion for mmWave radar platform vibratory motion compensation [1]. The chip analog front-end (AFE) interfaces directly with an external analog IMU to capture the parasitic motion sensed by an automotive FMCW radar, with circuits designed to detect the frequency and amplitude of the platform vibration. Digitized vibration parameter information is subsequently processed by an on-chip, custom DSP engine responsible for formulating the deconvolution kernel filter capable of inverting, and thereby correcting, the effect of the parasitic vibration. The DSP engine stores the deconvolution kernel phase and magnitude data in an on-chip SRAM bank for serial readout by the mmWave radar interface controller.

Measurements on the fabricated chip illustrate accurate generation of the appropriate motion deconvolution kernel in response to vibration signals of varying amplitude and frequency and exhibit impressive correction results when applied to simultaneously collected radar data, with low power consumption. The holistic mixed-signal chip, therefore, serves as a baseline proof-of-concept IC for real-time, fully autonomous on-chip motion compensation, a critical feature for high-resolution perception in intelligent edge sensors.

Adaptive AI-in-the-Loop mmWave Radar Transceiver

Currently, we are working on a fully adaptive mmWave transceiver chip with real-time adaptable front-end circuits for energy-efficient and high-resolution ML-based perception tasks. The chip is designed to interface with an off-chip ML engine, which takes as inputs the pre-processed ADC data translated on chip and outputs a set of tuning knobs shown to optimize the IC’s resource consumption and the resulting data size as a function of target class. This system-level AI-enabled feedback optimizes the range/Doppler sampling schemes and RF/analog front-end performance in accordance with the scene's detected range/Doppler characteristics to minimize power consumption, latency, and data size given a target classification accuracy or confidence. Ultimately, through this closed-loop dynamic framework, we aim to achieve a 100x performance/efficiency improvement over existing traditional open-loop edge perception architectures and commercial devices.

Publications

[1] N. Poole and A. Arbabian, “A 130-nm Fusion-Based Deconvolution Kernel Generator IC For Real-Time mmWave Radar Motion Compensation,” IEEE Access, vol. 11, Nov. 2023, pp. 132223-132238.