Real-time object detection using FPGA accelerators and OpenVINO toolkit

Abstract

Real-time object detection plays a crucial role in diverse applications, including autonomous vehicles, surveillance systems, and industrial automation. However, achieving real-time performance with deep learning models remains challenging due to the computational complexity and energy demands of traditional hardware platforms like GPUs. This study investigates the potential of Field-Programmable Gate Arrays (FPGAs) combined with the OpenVINO toolkit to optimize and accelerate real-time object detection tasks using YOLOv4 and SSD models. The primary objective was to evaluate the performance of FPGA accelerators in terms of inference latency, throughput, energy efficiency, and detection accuracy, comparing them with traditional GPU-based systems. The FPGA implementation was developed using an Intel Arria 10 FPGA board and optimized via the OpenVINO toolkit, while image acquisition was conducted using high-resolution sensors. Comparative evaluations were performed across the COCO and Pascal VOC datasets. The results demonstrated that FPGA significantly outperformed GPUs in key performance metrics. FPGA achieved lower inference latency (12.8 ms for YOLOv4, 10.2 ms for SSD) and higher throughput (78 FPS for YOLOv4, 85 FPS for SSD) compared to GPUs (18.5 ms latency and 65 FPS for YOLOv4; 14.9 ms latency and 72 FPS for SSD). Furthermore, FPGA showcased superior energy efficiency (1.95 FPS/W for YOLOv4, 2.12 FPS/W for SSD) compared to GPUs (0.31 FPS/W and 0.34 FPS/W, respectively). Detection accuracy remained consistent across platforms, with FPGA achieving comparable mean Average Precision (mAP) values. These findings underscore the advantages of FPGA accelerators for real-time object detection, particularly in edge computing environments. Future research should focus on addressing FPGA programming complexities, expanding hardware compatibility, and exploring hybrid FPGA-GPU architectures to optimize AI workloads further.