Scholarlink

This work presents a novel implementation of wideband Direction of Arrival (DOA) estimation using a Deep Learning approach deployed on an FPGA. A 4-element uniform linear antenna array acquires chirp-modulated signals, which are preprocessed and fed into a modified ResNet-18 convolutional neural network. The network performs angle regression across a ±60 \(^\circ \) range and achieves a Root Mean Square Error (RMSE) of 0.4175 \(^\circ \) . Deployment is carried out on the AMD ZCU102 board using MathWorks’ Deep Learning Processor IP with INT8 quantization, resulting in a processing time of 3.231 ms and dynamic power consumption of 12.376 W. This work shows the feasibility of AI on reconfigurable hardware for real-time WideBand DOA estimation, as a competitive alternative to traditional methods.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AI-Based WideBand Direction of Arrival Estimation on FPGA

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Lorenzo D’Antonio,
  • Sergio Spanò,
  • Riccardo La Cesa,
  • Cristian Valenti,
  • Udhaya Mugil Damodarin,
  • Luca Di Nunzio

摘要

This work presents a novel implementation of wideband Direction of Arrival (DOA) estimation using a Deep Learning approach deployed on an FPGA. A 4-element uniform linear antenna array acquires chirp-modulated signals, which are preprocessed and fed into a modified ResNet-18 convolutional neural network. The network performs angle regression across a ±60 \(^\circ \) range and achieves a Root Mean Square Error (RMSE) of 0.4175 \(^\circ \) . Deployment is carried out on the AMD ZCU102 board using MathWorks’ Deep Learning Processor IP with INT8 quantization, resulting in a processing time of 3.231 ms and dynamic power consumption of 12.376 W. This work shows the feasibility of AI on reconfigurable hardware for real-time WideBand DOA estimation, as a competitive alternative to traditional methods.