International Journal of Electronic Devices and Networking
2024, Vol. 5, Issue 2, Part A
Implementation of real-time object detection on Intel arria FPGA using pyTorch
Author(s): Ali Ben Amara, Salma Bouazizi and Mariem Jebali
Abstract: Real-time object detection has become a cornerstone in various fields, including autonomous systems, surveillance, and robotics. However, deploying computationally intensive deep learning models on resource-constrained hardware remains challenging. This study aims to implement and optimize real-time object detection models, YOLOv4 and Faster R-CNN, on an Intel Arria FPGA using PyTorch, focusing on improving latency, power efficiency, and frame-per-second (FPS) performance without significant loss of accuracy. The models were first trained and optimized in PyTorch using the COCO dataset, followed by quantization into an 8-bit fixed-point representation. OpenCL kernels were developed to manage FPGA resource utilization, memory allocation, and computational parallelism. The quantized models were compiled and deployed on the Intel Arria FPGA platform, and real-time inference performance was evaluated. Results revealed that YOLOv4 outperformed Faster R-CNN, achieving a latency of 18ms, FPS of 55, and power consumption of 35W, compared to 26ms latency, FPS of 38, and 40W power consumption for Faster R-CNN. The accuracy remained within acceptable limits after quantization, with YOLOv4 showing a slight reduction of ~1.3% and Faster R-CNN ~1.8%. Resource utilization analysis indicated that YOLOv4 consumed 82% of Look-Up Tables (LUTs) and 80% of DSPs, while Faster R-CNN consumed 74% LUTs and 80% DSPs. Statistical analysis using a paired t-test confirmed significant improvements (p < 0.05) in latency and FPS for FPGA deployments compared to GPUs. The study highlights the efficiency of Intel Arria FPGA in accelerating deep learning workloads and suggests future research on advanced quantization strategies, dynamic resource allocation, and enhanced PyTorch-to-FPGA compiler workflows. This research provides a scalable and energy-efficient framework for deploying object detection models on FPGA platforms.
DOI: 10.22271/27084477.2024.v5.i2a.61
Pages: 18-23 | Views: 44 | Downloads: 16
Download Full Article: Click Here
How to cite this article:
Ali Ben Amara, Salma Bouazizi, Mariem Jebali. Implementation of real-time object detection on Intel arria FPGA using pyTorch. Int J Electron Devices Networking 2024;5(2):18-23. DOI: 10.22271/27084477.2024.v5.i2a.61
