High-throughput perception-aware mmWave radar object recognition system

In recent years, millimeter-wave (mmWave) radar has garnered increasing attention. It has demonstrated strong performance under challenging visibility conditions such as fog, smoke, and darkness, challenging traditional cameras and LiDAR. Meanwhile, mmWave radar also supports high-resolution imaging through techniques such as synthetic aperture radar (SAR), which enables object recognition using computer vision models. The increasing availability of compact, low-cost mmWave systems has made them particularly appealing for low-power platforms in mobile scenarios. However, mainstream mmWave imaging solutions usually demand extensive sampling to achieve high resolution. This leads to significant delays and excessive data collection, particularly in environments where only a small subset of the scene contains target-relevant information. The high-resolution image fed to vision models also causes runtime overhead due to pixels describing an empty background. This intensive sampling requirement may lead to additional safety risks in time-critical scenarios like autonomous driving. Therefore, this thesis focuses on the question: How can intelligent processing and transmission of mmWave radar signals be used to reduce time and resource consumption in areas with low informational value? By analyzing the source and effects of these areas in a standard scanning-imaging-recognizing workflow, we identified three categories of data generated during the standard mmWave recognition workflow, which delay processing and cause low recognition output. For each type of data with low informational value, we designed a module to exclude data that contains a low informational value. All modules are integrated into mmPrism, a perception-aware attention scheduling system. By combining rapid coarse scanning with high-resolution SAR imaging under the control of a perception-aware region of interest (ROI) scheduler, this system minimizes time and resource consumption in areas with low informational value, answering the core question of this thesis. We conducted radar imaging and recognition acceleration experiments to evaluate our solution. Radar imaging testbeds are built using Texas Instruments’ IWR mmWave radar, proving the capabilities of a fast full-scene sweeping and region-selective high-resolution imaging. By excluding data with low informational value, mmPrism improves recognition throughput by up to 47× and reduces end-to-end latency by over 90% for scenes exceeding 60 meters in lateral extent. Sampling volume and processing overhead also decrease quadratically. This substantial improvement shows the potential of our system design to be a solution for next-generation mmWave-based recognition systems.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Shan, Zihan

Contributors dc:contributor

• Caesar, Matthew

• Mmwave
• Cognitive Radar
• Automatic Target Recognition
• Internet Of Things

Statement dc:rights

• Copyright 2025 Zihan Shan

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/130165

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F130165