Page 342 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

Security & Defense

The solution strengthens urban safety and national security by providing real-time, energy-
efficient RF-based surveillance for intrusion detection, border monitoring, and critical
infrastructure protection.

Cost-Effective and Scalable Deployment

AI-driven RF sensing reduces infrastructure costs while increasing energy efficiency and
scalability. It enables resilient smart city systems that are easier to deploy and manage across
diverse urban environments.

2�3 Future Work

• Large-scale RF Dataset Development: Building extensive, diverse RF datasets is essential
to improve AI model generalization. Future work should focus on creating open,
standardized RF datasets covering various environments (urban, rural, indoor, outdoor),
object types, materials, and propagation conditions (LoS/NLoS).
• Optimizing AI Model Complexity: Reducing the computational load of deep learning
models without compromising accuracy remains a key challenge. Future research
should explore model compression techniques (e.g., pruning, quantization, knowledge
distillation) and lightweight architectures optimized for real-time RF sensing.
• Seamless Integration with Low-Latency 5G Processing: Future systems must ensure tight
integration of AI-driven object detection with 5G baseband pipelines. This includes
designing low-overhead interfaces between AI blocks and physical layer processing,
and leveraging AI for real-time beamforming, interference management, and dynamic
spectrum access in ISAC scenarios.

3 Use Case Requirements

• REQ-01: It is expected that the system supports real-time RF-based object detection with
latency less than or equal to 1 millisecond.
• REQ-02: It is expected that field-programmable gate array (FPGA)-based artificial
intelligence (AI) modules are utilized for noise filtering, interference mitigation, and signal
clarity enhancement.
• REQ-03: It is critical that the system accurately extracts radio frequency (RF) features
including time delay, angle of arrival (AoA), and signal strength to enable precise object
characterization.
• REQ-04: It is expected that edge inference using convolutional neural networks (CNNs)
achieves a throughput of at least 100 inferences per second when accelerated with tensor
processing hardware.
• REQ-05: It is critical that the system maintains robust classification performance in both
line-of-sight (LoS) and non-line-of-sight (NLoS) environments, across urban, rural, and
low-visibility scenarios.
• REQ-06: It is expected that the solution is fully compatible with 5G New Radio (5G-NR)
basebands to enable integrated sensing and communication (ISAC) deployment.
• REQ-07: It is critical that training and validation are performed using structured RF
datasets that include both controlled anechoic environments and real-world deployment
scenarios.

306