YOMANet-Accel:面向边缘端人车检测的轻量化算法加速器

陈宁江; 卢耀宗

doi:10.11999/JEIT250059

YOMANet-Accel:面向边缘端人车检测的轻量化算法加速器

doi: 10.11999/JEIT250059 cstr: 32379.14.JEIT250059

陈宁江^{1, 2, 3},
卢耀宗^1, ,

1.
广西大学计算机与电子信息学院南宁 530004
2.
广西智能数字服务技术创新中心南宁 530004
3.
广西高校并行分布与智能计算重点实验室南宁 530004

基金项目: 国家自然科学基金(62162003)，中央引导地方科技发展资金(桂科ZY24212059)

详细信息

作者简介:
陈宁江：男，博士，教授，研究方向为智能软件工程、云计算、大数据、分布式计算、边缘计算

卢耀宗：男，硕士生，研究方向为边缘计算

通讯作者:
卢耀宗　luyaozong9725@163.com

中图分类号: TN911.7; TN4; TP389.1
计量
- 文章访问数: 206
- HTML全文浏览量: 96
- PDF下载量: 40
- 被引次数: 0
出版历程
- 收稿日期: 2025-01-22
- 修回日期: 2025-06-30
- 网络出版日期: 2025-07-04
- 刊出日期: 2025-08-27

YOMANet-Accel: A Lightweight Algorithm Accelerator for Pedestrians and Vehicles Detection at the Edge

CHEN Ningjiang^{1, 2, 3},
LU Yaozong^{1
, ,}

1.
School of Computer, Electronics and Information, Guangxi University, Nanning 530004, China
2.
Guangxi Center of Technology Innovation for Intelligent Digital Services, Nanning 530004, China
3.
Key Laboratory of Parallel, Distributed and Intelligent Computing (Guangxi University), Education Department of Guangxi Zhuang Autonomous Region, Nanning 530004, China

Funds: The National Natural Science Foundation of China (62162003), The Central Guidance on Local Science and Technology Development Fund of Guangxi Province (GuikeZY24212059)

摘要

摘要: 针对自动驾驶边缘计算场景中行人车辆检测任务面临的模型计算复杂度高、参数量大导致的部署难题，该文提出一种轻量化神经网络模型YOMANet (Yolo Model Adaptation Network)，基于异构FPGA平台设计YOMANet加速器(YOMANet-Accel)，实现边缘端人车检测的算法加速。YOMANet算法的主干网络采用轻量型网络MobileNetv2以大幅压缩模型参数量，颈部网络使用深度可分离卷积来代替常规卷积以提升训练速度，并在头部网络嵌入基于归一化的注意力模块(NAM)以增强网络对细节信息的捕获能力。为将YOMANet算法部署到现场可编程门阵列(FPGA)平台，该文针对卷积运算在任务层设计循环分块以调整内循环和外循环的顺序，在运算层对处理引擎单元(PE)设计乘加树，使得多个乘加运算可以同时执行，提高数据的并行计算效率。同时，针对数据存储过程采用双缓存机制来减少数据传输时延，对权重参数和激活函数进行int8数据量化以降低资源消耗。实验结果表明，YOMANet算法在训练平台上的检测精度和检测速度表现优异，对小目标和遮挡目标具备较好的检测能力，有效减少了误检和漏检情况的发生。算法部署到硬件平台后，YOMANet-Accel的目标检测效果保持在较高水平，硬件资源的能效比表现良好，有效发挥了FPGA的并行优势。
- 行人与车辆检测 /
- 边缘计算 /
- 轻量化 /
- 异构FPGA加速
Abstract: Objective Accurate and real-time detection of pedestrians and vehicles is essential for autonomous driving at the edge. However, deep learning-based object detection algorithms are often challenging to deploy in edge environments due to their high computational demands and complex parameter structures. To address these limitations, this study proposes a soft-hard coordination strategy. A lightweight neural network model, Yolo Model Adaptation Network (YOMANet), is designed, and a corresponding neural network accelerator, YOMANet Accelerator (YOMANet-Accel), is implemented on a heterogeneous Field-Programmable Gate Array (FPGA) platform. This system enables efficient algorithm acceleration for pedestrian and vehicle detection in edge-based autonomous driving scenarios. Methods The lightweight backbone of YOMANet adopts MobileNetv2 to reduce the number of network parameters. The neck network incorporates the Spatial Pyramid Pooling (SPP) and Path Aggregation Network (PANet) structures from YOLOv4 to expand the receptive field and accommodate targets of varying sizes. Depthwise separable convolution replaces standard convolution, thereby reducing training complexity and improving convergence speed. To enhance detail extraction, the Normalization-based Attention Module (NAM) is integrated into the head network, allowing suppression of irrelevant feature weights. For deployment on a FPGA platform, parallel computing and data storage schemes are designed. The parallel computing strategy adopts a loop blocking method to reorder inner and outer loops, enabling access to different output array elements through adjacent loop layers and facilitating parallel processing of output feature map pixels. Multiply-add trees are implemented in the Processing Engine (PE) to support efficient task allocation and operation scheduling. A double-buffer mechanism is introduced in the data storage scheme to increase data reuse, minimize transmission latency, and enhance system throughput. In addition, int8 quantization is applied to both weight parameters and activation functions, reducing the overall parameter size and accelerating parallel computation. Results and Discussions Experimental results on the training platform indicate that YOMANet achieves the inference speed characteristic of lightweight models while maintaining the detection accuracy of large-scale models, thereby improving overall detection performance (Fig. 12, Table 2). The ablation study demonstrates that the integration of MobileNetv2 and depthwise separable convolution significantly reduces the number of model parameters. Embedding the NAM attention mechanism does not noticeably increase model size but enhances detail extraction and improves detection of small targets (Table 3). Compared with other lightweight algorithms, the enhanced YOMANet shows improved detail extraction and superior detection of small and occluded targets, with substantially lower false and missed detection rates (Fig. 13). Results on the accelerator platform reveal that quantization has minimal effect on accuracy while substantially reducing model size, supporting deployment on resource-constrained edge devices (Table 4). When deployed on the FPGA platform, YOMANet retains detection accuracy comparable to GPU/CPU platforms, while power consumption is reduced by an order of magnitude, meeting the efficiency requirements for edge deployment (Fig. 14). Compared with related accelerator designs, YOMANet-Accel achieves competitive throughput and the highest Digital Signal Processing (DSP) efficiency, demonstrating the effectiveness of the proposed parallel computing and storage schemes in utilizing FPGA resources (Table 5). Conclusions Experimental results demonstrate that YOMANet achieves high detection accuracy and fast inference speed on the training platform, with enhanced performance for small and occluded targets, leading to a reduced missed detection rate. When deployed on the FPGA platform, YOMANet-Accel achieves an effective balance between detection performance and resource efficiency, supporting real-time pedestrian and vehicle detection in edge computing scenarios.
- Pedestrians and vehicles detection /
- Edge computing /
- Lightweight model /
- Heterogeneous FPGA acceleration

HTML全文

图 1 YOMANet轻量化神经网络模型的架构

下载: 全尺寸图片幻灯片

图 2 Bottleneck模块的反向残差结构

下载: 全尺寸图片幻灯片

图 3 标准卷积过程

下载: 全尺寸图片幻灯片

图 4 深度可分离卷积过程

下载: 全尺寸图片幻灯片

图 5 通道注意力子模块示意图

下载: 全尺寸图片幻灯片

图 6 空间注意力子模块示意图

下载: 全尺寸图片幻灯片

图 7 YOMANet-Accel整体架构

下载: 全尺寸图片幻灯片

图 8 循环分块技术

下载: 全尺寸图片幻灯片

图 9 PE乘加树设计

下载: 全尺寸图片幻灯片

图 10 双缓存机制

下载: 全尺寸图片幻灯片

图 11 目标数据集各类别数量占比

下载: 全尺寸图片幻灯片

图 12 目标类别的Precision, Recall以及AP值

下载: 全尺寸图片幻灯片

图 13 不同轻量化算法的图像检测效果对比

下载: 全尺寸图片幻灯片

图 14 YOMANet算法在不同平台上的性能表现

下载: 全尺寸图片幻灯片

表 1 YOMANet主干网络模型结构

Input	Operate	t	c	S	Input	Operate	t	c	S
416×416×3	Conv2d	-	32	2	26×26×32	Bottleneck5-4	6	64	1
208×208×32	Bottleneck2-1	1	16	1	26×26×64	Bottleneck6-1	6	64	1
208×208×16	Bottleneck3-1	6	16	2	26×26×64	Bottleneck6-2	6	64	1
104×104×16	Bottleneck3-2	6	24	1	26×26×64	Bottleneck6-3	6	96	1
104×104×24	Bottleneck4-1	6	24	2	26×26×96	Bottleneck7-1	6	96	2
52×52×24	Bottleneck4-2	6	24	1	13×13×96	Bottleneck7-2	6	96	1
52×52×24	Bottleneck4-3	6	32	1	13×13×96	Bottleneck7-3	6	160	1
52×52×32	Bottleneck5-1	6	32	2	13×13×160	Bottleneck8-1	6	320	1
26×26×32	Bottleneck5-2	6	32	1	13×13×320	DSConv×3	-	-	-
26×26×32	Bottleneck5-3	6	32	1

下载: 导出CSV

表 2 不同算法在GPU平台上的性能比较

Model	Backbone	Input size	Data type	mAP@0.5(%)	Size(MB)	fps
Faster RCNN	VGG16	600×1000	float32	90.13	500.67	18
SSD	VGG16	512×512	float32	88.13	287.64	28
YOLOv4	CSP-DarkNet53	416×416	float32	88.29	249.48	39
YOLOv5	CSP-DarkNet53	640×640	float32	88.72	223.62	51
CCBA-NMS-YD^[7]	VGG16	512×512	float32	87.06	-	-
YOLOv3-Improved^[8]	Darknet53	416×416	float32	86.24	-	-
YOLOv5s-RFB-s-ASFF^[9]	CSP-RFB-s-ASFF	640×640	float32	84.01	-	61
YOLOP-E^[28]	EfficientNetv2	-	float32	79.20	27.6	41.6
YOLOv4-tiny	CSP-Darknet53-tiny	416×416	float32	80.69	37.94	74
YOLOv5s	CSP-DarkNet53	640×640	float32	83.51	34.46	78
YOLOv7-tiny	CSP-PANet	416×416	float32	86.68	32.75	84
YOMANet	MobileNetv2	416×416	float32	88.26	30.95	80

下载: 导出CSV

表 3 消融实验性能比较

MobileNetv2	DSConv	NAM	Data type	mAP(%)	Size(MB)	fps	Power(W)
×	×	×	float32	88.29	249.48	39	167.436
√	×	×	float32	86.76	80.24	63	123.481
√	√	×	float32	86.37	28.89	84	106.274
√	√	√	float32	88.26	30.95	80	108.915

下载: 导出CSV

表 4 数据量化前后效果对比

Model	Data type	mAP@0.5(%)	Size(MB)
YOMANet	float32	88.26	30.95
YOMANet	int8	86.23	7.76

下载: 导出CSV

表 5 与相关文献的加速器性能对比

	文献[11]	文献[12]	文献[13]	文献[14]	文献[29]	本文
Model	YOLOv4-tiny	YOLOv3-tiny	YOLOv5s	Ultranet	YOLOv3-tiny	YOMANet
Backbone	CSP-Darknet53	DarkNet53	CSP-DarkNet53	VGG16	DarkNet53	MobileNetv2
FPGA	ZYNQ 7020	Nexys A7-100T	ZYNQ 7020	ZYNQ ZU3EG	ZYNQ XCZU9EG	ZYNQ 7020
DSP	220	240	220	360	298	220
Data type	16 bit	int8	16 bit	4 bit	16 bit	int8
Power (W)	2.750	2.203	3.039	6.650	4.120	7.402
GOPS	-	95.08	30.10	126.72	96.60	100.23
GOPS/DSP	-	0.396	0.137	0.352	0.324	0.456
GOPS/W	-	43.16	9.90	19.06	23.45	13.54
Size (MB)	23.70	-	-	15.18	-	7.76
fps (帧/s)	-	76.75	-	220.76	17.30	40.15
mAP@0.5(%)	77.80	81.19	40.30	-	31.50	86.23

下载: 导出CSV

参考文献(29)

[1]	ZHAN Jiao, LIU Jingnan, WU Yejun, et al. Multi-task visual perception for object detection and semantic segmentation in intelligent driving[J]. Remote Sensing, 2024, 16(10): 1774. doi: 10.3390/rs16101774.
[2]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 779–788. doi: 10.1109/CVPR.2016.91.
[3]	谭郁松, 李恬, 张钰森. 面向边缘智能的神经网络模型生成与部署研究[J]. 计算机工程, 2024, 50(8): 1–12. doi: 10.19678/j.issn.1000-3428.0068554. TAN Yusong, LI Tian, and ZHANG Yusen. Research on neural network model generation and deployment for edge intelligence[J]. Computer Engineering, 2024, 50(8): 1–12. doi: 10.19678/j.issn.1000-3428.0068554.
[4]	REN Shaoqing, HE Kaiming, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137–1149. doi: 10.1109/TPAMI.2016.2577031.
[5]	LIU Wei, ANGUELOV D, ERHAN D, et al. SSD: Single shot MultiBox detector[C]. The 14th European Conference on Computer Vision, Amsterdam, The Netherlands, 2016: 21–37. doi: 10.1007/978-3-319-46448-0_2.
[6]	WEI Hongyang, ZHANG Qianqian, HAN Jingjing, et al. SARNet: Spatial Attention Residual Network for pedestrian and vehicle detection in large scenes[J]. Applied Intelligence, 2022, 52(15): 17718–17733. doi: 10.1007/s10489-022-03217-9.
[7]	YUAN Zhenhao, WANG Zhiwen, and ZHANG Ruonan. CCBA-NMS-YD: A Vehicle pedestrian detection and tracking method based on improved YOLOv7 and DeepSort[J]. World Electric Vehicle Journal, 2024, 15(7): 309. doi: 10.3390/wevj15070309.
[8]	王启明, 何梓林, 张栋林, 等. 基于YOLOv3的雾天场景行人车辆检测方法研究[J]. 控制工程, 2024, 31(3): 510–517. doi: 10.14107/j.cnki.kzgc.20211118. WANG Qiming, HE Zilin, ZHANG Donglin, et al. Research on pedestrian and vehicle detection method based on YOLOv3 in foggy scene[J]. Control Engineering of China, 2024, 31(3): 510–517. doi: 10.14107/j.cnki.kzgc.20211118.
[9]	胡丹丹, 张忠婷. 基于改进YOLOv5s的面向自动驾驶场景的道路目标检测算法[J]. 智能系统学报, 2024, 19(3): 653–660. doi: 10.11992/tis.202206034. HU Dandan and ZHANG Zhongting. Road target detection algorithm for autonomous driving scenarios based on improved YOLOv5s[J]. CAAI Transactions on Intelligent Systems, 2024, 19(3): 653–660. doi: 10.11992/tis.202206034.
[10]	WANG Haotian, ZHAO Yinghai, and GAO Fan. A convolutional neural network accelerator based on FPGA for buffer optimization[C]. 2021 IEEE 5th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC), Chongqing, China, 2021: 2362–2367. doi: 10.1109/IAEAC50856.2021.9390606.
[11]	ZHAO Sijie, GAO Shangshang, WANG Rugang, et al. Acceleration and implementation of convolutional neural networks based on FPGA[J]. Digital Signal Processing, 2023, 141: 104188. doi: 10.1016/j.dsp.2023.104188.
[12]	KIM M, OH K, CHO Y, et al. A low-latency FPGA accelerator for YOLOv3-tiny with flexible layerwise mapping and dataflow[J]. IEEE Transactions on Circuits and Systems I: Regular Papers, 2024, 71(3): 1158–1171. doi: 10.1109/TCSI.2023.3335949.
[13]	刘谦, 王林林, 周文勃. 基于FPGA的YOLOv5s网络高效卷积加速器设计[J]. 电讯技术, 2024, 64(3): 366–375. doi: 10.20079/j.issn.1001-893x.230216003. LIU Qian, WANG Linlin, and ZHOU Wenbo. Design of a YOLOv5s network efficient convolution accelerator powered by FPGA[J]. Telecommunication Engineering, 2024, 64(3): 366–375. doi: 10.20079/j.issn.1001-893x.230216003.
[14]	包振山, 郭俊南, 张文博, 等. UltraAcc: 基于FPGA流水架构的低功耗高性能CNN加速器定制设计[J]. 计算机学报, 2023, 46(6): 1139–1155. doi: 10.11897/SP.J.1016.2023.01139. BAO Zhenshan, GUO Junnan, ZHANG Wenbo, et al. UltraAcc: A customized low power and high performance CNN accelerator with dataflow on FPGAs[J]. Chinese Journal of Computers, 2023, 46(6): 1139–1155. doi: 10.11897/SP.J.1016.2023.01139.
[15]	WANG C Y, BOCHKOVSKIY A, and LIAO H Y M. Scaled-YOLOv4: Scaling cross stage partial network[C]. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, USA, 2021: 13024–13033. doi: 10.1109/CVPR46437.2021.01283.
[16]	JOCHER G, STOKEN A, BOROVEC J, et al. ultralytics/YOLOv5: V4.0 - nn. SiLU() activations, Weights & Biases logging, PyTorch Hub integration[Z]. 2021. doi: 10.5281/ZENODO.4418161.
[17]	WANG C Y, BOCHKOVSKIY A, and LIAO H Y M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, Canada, 2022: 7464–7475. doi: 10.1109/CVPR52729.2023.00721.
[18]	SANDLER M, HOWARD A, ZHU Menglong, et al. MobileNetV2: Inverted residuals and linear bottlenecks[C]. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 4510–4520. doi: 10.1109/CVPR.2018.00474.
[19]	LIU Yichao, SHAO Zongru, TENG Yueyang, et al. NAM: Normalization-based attention module[EB/OL]. https://arxiv.org/abs/2111.12419, 2021.
[20]	ZHANG Xiangyu, ZHOU Xinyu, LIN Mengxiao, et al. ShuffleNet: An extremely efficient convolutional neural network for mobile devices[C]. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 6848–6856. doi: 10.1109/CVPR.2018.00716.
[21]	TAN Mingxing and LE Q. EfficientNet: Rethinking model scaling for convolutional neural networks[C]. The 36th International Conference on Machine Learning, Long Beach, USA, 2019: 6105–6114.
[22]	HAN Kai, WANG Yunhe, TIAN Qi, et al. GhostNet: More features from cheap operations[C]. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, USA, 2020: 1577–1586. doi: 10.1109/CVPR42600.2020.00165.
[23]	HOWARD A G, ZHU Menglong, CHEN Bo, et al. MobileNets: Efficient convolutional neural networks for mobile vision applications[EB/OL]. https://arxiv.org/abs/1704.04861, 2017.
[24]	HU Jie, SHEN Li, SUN Gang, et al. Squeeze-and-excitation networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(8): 2011–2023. doi: 10.1109/TPAMI.2019.2913372.
[25]	WANG Qilong, WU Banggu, ZHU Pengfei, et al. ECA-Net: Efficient channel attention for deep convolutional neural networks[C]. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, USA, 2020: 11531–11539. doi: 10.1109/CVPR42600.2020.01155.
[26]	WOO S, PARK J, LEE J Y, et al. CBAM: Convolutional block attention module[C]. The 15th European Conference on Computer Vision, Munich, Germany, 2018: 3–19. doi: 10.1007/978-3-030-01234-2_1.
[27]	CORDTS M, OMRAN M, RAMOS S, et al. The cityscapes dataset for semantic urban scene understanding[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 3213–3223. doi: 10.1109/CVPR.2016.350.
[28]	LIU Yulin, LI Gang, HAO Liguo, et al. Research on a lightweight panoramic perception algorithm for electric autonomous mini-buses[J]. World Electric Vehicle Journal, 2023, 14(7): 179. doi: 10.3390/wevj14070179.
[29]	任仕伟, 刘朝钾, 李剑铮, 等. 面向端到端目标检测神经网络的高效硬件加速系统设计[J]. 北京理工大学学报, 2022, 42(12): 1312–1320. doi: 10.15918/j.tbit1001-0645.2022.004. REN Shiwei, LIU Chaojia, LI Jianzheng, et al. Efficient hardware acceleration system design for end-to-end object detection neural network[J]. Transactions of Beijing Institute of Technology, 2022, 42(12): 1312–1320. doi: 10.15918/j.tbit1001-0645.2022.004.