Application of Deep Learning on Millimeter-Wave Radar Signals: A Review

Area of Science:

• Artificial Intelligence
• Computer Vision
• Robotics

Background:

• Deep learning (DL) models predominantly use Lidar or camera data, overlooking automotive radar's potential.
• Automotive radar offers unique advantages in adverse weather and simultaneous range/velocity measurement.
• Limited benchmark data has historically hindered radar's application in DL, but this is changing with new datasets.

Purpose of the Study:

• To survey deep learning approaches for processing automotive radar signals in autonomous driving.
• To review deep learning-based multi-sensor fusion models using radar and camera data for object detection.
• To summarize available radar datasets and identify research gaps and future prospects.

Main Methods:

• Systematic review of deep learning techniques applied to radar signal processing.
• Categorization of methods based on radar signal representations.
• Analysis of multi-sensor fusion models integrating radar and camera data.
• Compilation of existing radar datasets for autonomous driving.

Main Results:

• Deep learning models are being increasingly applied to radar data for object detection and classification.
• Multi-sensor fusion approaches combining radar and camera data show significant promise.
• Several new datasets are emerging, facilitating further research in this domain.

Conclusions:

• Automotive radar, despite its potential, remains under-explored in deep learning applications.
• Standardizing radar data representation and developing more benchmark datasets are crucial.
• Future research should focus on advanced fusion techniques and robust radar-based perception systems.