电子器件2025，Vol.48Issue(2)：285-291,7.DOI:10.3969/j.issn.1005-9490.2025.02.008
高能效目标检测硬件加速器设计
High-Efficiency Hardware Accelerator Design for Object Detection
摘要
Abstract
Existing hardware acceleration solutions for object detection algorithms suffer from poor portability,high power consumption,and large resource overhead.Focusing on the lightweight and fast YOLO-FastestV2 algorithm,a high-efficiency and low-overhead accel-erator for object detection is designed.Firstly,YOLO Fastest-V2 is sequentially pruned,trained and quantized.Secondly,the data access methods and pipeline design are optimized.In addition,a dynamic design strategy for convolutional parallelism is proposed to maximize resource utilization efficiency.Finally,the proposed accelerator is verified on ZYNQ-MPSOC platform.The experimental results demon-strate that the detection speed of the proposed accelerator reaches 147 FPS,while consuming 0.995 W of power,144 DSPs and 125 BRAM36kb.The energy efficiency ratio of the accelerator is 147.74 FPS/W,which is higher than the reported hardware accelera-tors for object detection.关键词
目标检测/硬件加速/ZYNQ/YOLO Fastest-V2Key words
object detection/hardware acceleration/ZYNQ/YOLO Fastest-V2分类
电子信息工程引用本文复制引用
谢玉东,林枫,霍亮州,申志浩,陈鑫,张颖..高能效目标检测硬件加速器设计[J].电子器件,2025,48(2):285-291,7.基金项目
国家自然科学基金项目(61106029) （61106029）
国防科技重点实验室基金项目(61425010302) （61425010302）
中国航空发动机集团产学研合作项目(HFZL2022CXY012) （HFZL2022CXY012）
