Page 339 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Testbeds or Pilot N/A
Deployments 4.3 - 5G
Code repositories N/A

2 Use Case Description

2�1 Description

AI-driven RF-based object detection and classification for 5G-Advanced and beyond leverages
Integrated Sensing and Communication (ISAC) to enable real-time monitoring and security
applications across smart transportation, industrial automation, border surveillance, and
next-generation telecommunications. As 5G technology progresses toward 5G-Advanced,
integrating object detection and classification into the 5G New Radio (5G-NR) baseband
processing framework becomes essential for critical applications. This includes pedestrian
safety, where smart transportation systems require real-time object detection to prevent
accidents, intruder detection for protecting infrastructure, and border security, where ISAC
enhances surveillance and threat detection.

Unlike conventional vision-based detection, which relies on structured image data, RF sensing
processes reflected wireless signals that are non-independent and identically distributed
(non-IID) due to multipath propagation, interference, and material properties. AI is crucial
in transforming RF reflections into structured correlated data for object representation by
addressing challenges such as nonlinear distortions, multipath interference, and environmental
variations.

The AI processing pipeline begins at the baseband layer, where FPGA-based AI models filter
noise, mitigate multipath distortions, and enhance signal clarity. Feature extraction is performed
by analyzing delay spread, angle of arrival (AoA), and signal strength variations to create multi-
dimensional feature maps for object characterization. CNNs trained on RF feature maps enable
real-time inference and mitigation via edge-based tensor processors. Additionally, temporal
AI models capture signal variations over time, improving classification accuracy for dynamic
objects. AI enables real-time adaptation to diverse conditions, ensuring robust detection in
low-visibility scenarios such as fog, darkness, and occlusions where optical sensors fail.

Edge detection in RF-based sensing differs significantly from traditional image processing.
While Sobel and Canny filters are commonly used for vision-based systems, RF-based
sensing requires alternative methods to define object boundaries. Signal strength variations
help estimate object shape, while multi-path reflections provide clues about object edges.
Additionally, RF sensing can determine whether an object is hollow or solid by analyzing the
reflection pattern—hollow objects produce two distinct edge reflections at the ends, while solid
objects generate continuous reflections across their structure. This information enhances the
accuracy of AI-driven object classification and environmental mapping.

303