FPGA-Powered Environment Awareness via Quantized Neural Networks for LiDAR-Aided mm-Wave Beam Prediction

Environment awareness can be highly beneficial for robust and agile beam prediction, particularly for beyond-5G millimeter-wave (mm-wave) networks. While machine learning (ML) algorithms have shown potential in leveraging external sensor data, such as LiDAR, radar, and cameras, to enhance beam prediction, existing solutions often rely on simulations or offline data processing, limiting their applicability in real-world deployments. Real-world deployment of such solutions requires low-latency ML inference tailored to resource-constrained hardware platforms but remains challenging due to the computational complexity and latency demands of ML models. This paper addresses these challenges by implementing real-time ML inference in hardware focusing on environment awareness using LiDAR data. We present an FPGA-based implementation of Quantized Neural Networks (QNNs) optimized for real-time LiDAR-aided beam prediction in beyond-5G mm-wave networks. Evaluations on the ZCU104 FPGA platform using real-world datasets demonstrate inference latencies in the tens to hundreds of microseconds, achieving comparable accuracy to state-of-the-art methods with only 2-bit weights and activations. Our results underline the effectiveness of the QNNs in achieving high accuracy, low latency, and hardware resource efficiency for real-world mm-wave applications.