2024, Vol. 5, Issue 2, Part A
Real-time object detection using Xilinx Zynq FPGA and tensor flow framework
Author(s): Amina El Mansouri, Hassan Ben Youssef and Leila Saidi
Abstract: Real-time object detection plays a pivotal role in fields such as autonomous systems, surveillance, robotics, and industrial automation. However, traditional hardware platforms, including CPUs and GPUs, often face limitations in latency, power efficiency, and scalability when deployed in resource-constrained environments. This study investigates the integration of Xilinx Zynq FPGA with the TensorFlow framework to address these challenges and enhance real-time object detection performance. The primary objectives were to design and implement an optimized object detection system using YOLOv3 on FPGA, evaluate its performance metrics, and compare them with CPU and GPU implementations. The YOLOv3 model was trained on the COCO dataset, quantized, and deployed onto the FPGA using Vivado HLS and PetaLinux OS for runtime execution. Key performance metrics, including frame rate (FPS), latency, power consumption, and mean Average Precision (mAP), were measured and statistically analyzed using ANOVA and paired t-tests. The results demonstrated that the FPGA implementation achieved an average frame rate of 35 FPS, latency of 28 ms, power consumption of 15 W, and mAP of 70.5%. While the FPGA marginally lagged behind GPU in detection accuracy (72.8%), it significantly outperformed CPU (65.4%) and exhibited superior power efficiency. Statistical validation confirmed significant differences across all performance metrics (p< 0.05). The findings emphasize FPGA’s suitability for low-latency, energy-efficient real-time applications despite minor trade-offs in accuracy. Practical recommendations include exploring dynamic reconfiguration, advanced quantization-aware training, hybrid FPGA-GPU architectures, and standardized deployment pipelines. This study concludes that Xilinx Zynq FPGA integrated with TensorFlow provides a scalable and efficient solution for real-time object detection, with potential applications in autonomous systems, edge computing, and smart surveillance technologies.
DOI: 10.22271/27084531.2024.v5.i2a.70
Pages: 13-18 | Views: 104 | Downloads: 42
Download Full Article: Click Here
How to cite this article:
Amina El Mansouri, Hassan Ben Youssef, Leila Saidi. Real-time object detection using Xilinx Zynq FPGA and tensor flow framework. Int J Res Circuits Devices Syst 2024;5(2):13-18. DOI: 10.22271/27084531.2024.v5.i2a.70
