《电子技术应用》
您所在的位置:首页 > 人工智能 > 设计应用 > 基于ZYNQ的YOLOv5n加速目标检测研究
基于ZYNQ的YOLOv5n加速目标检测研究
集成电路应用
张跃,张骢
沈阳理工大学装备工程学院
摘要: 针对嵌入式边缘计算场景中目标检测算法计算复杂度高与硬件资源受限的矛盾,该文选取参数轻量化的YOLOv5n网络结构,基于Xilinx ZYNQ7100异构平台设计专用加速器架构。采用量化感知训练方法优化网络模型,对优化后的模型执行INT8量化处理,设计乒乓双缓存机制提升数据传输效率,引入层间流水线并行策略提高计算并行度,基于Google TPU脉动阵列架构实现卷积计算硬件映射。研究表明,系统吞吐性能达到99 GOPS,平均推理速度达到22帧/秒,系统整体功耗为5.37 W,能效比达到18.44 GOPS/W。该方法在保证检测精度的同时,实现高能效目标检测,适用于资源受限的嵌入式边缘计算场景。
中图分类号:TP391文献标识码:ADOI:10.19339/j.issn.1674-2583.2026.06.006
中文引用格式:张跃,张骢. 基于ZYNQ的YOLOv5n加速目标检测研究[J].集成电路应用,2026,43(6):33-39.
英文引用格式:Zhang Yue,Zhang Cong. Research on YOLOv5n accelerated object detection based on ZYNQ[J].Application of IC,2026,43(6):33-39.
Research on YOLOv5n accelerated object detection based on ZYNQ
Zhang Yue,Zhang Cong
School of Equipment Engineering, Shenyang Ligong University
Abstract: To address the contradiction between high computational complexity of object detection algorithms and hardware resource constraints in embedded edge computing scenarios, a lightweight YOLOv5n network architecture is selected, and a dedicated accelerator architecture is designed based on the Xilinx ZYNQ7100 heterogeneous platform. Quantitative-aware training methods were used to optimize the network model, with INT8 quantization applied to the optimized model. A ping-pong double buffering mechanism was designed to improve data transmission efficiency; an interlayer pipeline parallel strategy was introduced to enhance computational parallelism; and hardware mapping for convolutional computing was implemented based on the Google TPU systolic array architecture. Research shows that the system throughput performance reached 99 GOPS, an average inference speed of 22 frames per second, overall system power consumption of 5.37 W, and the energy efficiency ratio of 18.44 GOPS/W. This method ensures detection accuracy while achieving high-efficiency object detection, suitable for resource-constrained embedded edge computing scenarios.
Key words : YOLOv5n; ZYNQ;object detection;hardware acceleration;quantitative-aware training

引言

随着人工智能与嵌入式视觉技术的快速发展,基于深度学习的目标探测与识别技术逐渐成为智能信息处理领域的重要研究方向。特别是单阶段检测框架YOLO系列算法以其简单的网络结构、检测速度快、便于工程部署等特点,广泛应用于智能监视、无人系统等领域。相比CPU、GPU和ASIC,FPGA能够利用片上DSP、BRAM等资源构建并行乘加计算单元,并通过流水线方式提高计算资源利用率。因此,将轻量化目标检测网络部署于FPGA平台,是实现嵌入式目标探测系统低延迟和低功耗计算的一种有效技术途径。

基于上述研究背景,本文以YOLOv5n通用目标检测网络为研究对象,面向嵌入式目标探测应用开展FPGA硬件加速设计。研究重点并非针对某一特定类别或特定尺度目标进行专门检测,而是围绕YOLOv5n网络整体推理过程中的计算和数据传输瓶颈,对模型量化、卷积计算映射、脉动阵列[1]、片上缓存以及数据传输机制进行协同设计。本文在ZYNQ7100平台上构建PSPL异构加速系统,将YOLOv5n的主要计算任务由软件处理器迁移至FPGA可编程逻辑部分,并结合DMA和乒乓双缓存机制(Pingpong Double Buffering)实现计算与数据传输的流水化执行。通过对加速器的功能正确性、资源利用率、推理性能以及功耗等指标进行测试与分析,验证所设计硬件架构的可行性和有效性。


本文详细内容请下载:

https://www.chinaaet.com/resource/share/2000007279


投稿邮箱:ICAPP@belling.com.cn


作者信息:

张跃,张骢

(沈阳理工大学装备工程学院,辽宁沈阳110159)

2.jpg

此内容为AET网站原创,未经授权禁止转载。