改进YOLOv4-tiny网络的狭小空间目标检测方法

doi:10.3778/j.issn.1002-8331.2112-0593

摘要/Abstract

摘要： 针对狭小空间中目标相互遮挡导致轻型检测网络存在大量漏检、分类错误等问题，基于YOLOv4-tiny提出一种自适应非极大抑制（adaptive non-maximum suppression，A-NMS）的多尺度检测方法。在骨干网络引入大尺度特征图优化策略和金字塔池化模型，增强遮挡目标显著区域特征；设计内嵌空间注意力的双路金字塔特征融合网络，提升浅层细节特征与高级语义信息的融合能力；提出区域目标密度与边界框中心距离因子相关联的动态NMS阈值设定方法，并在后处理阶段代替传统IoU-NMS算法，进一步减少漏检。实验结果表明，与YOLOv4-tiny算法相比，改进算法在公开数据集PASCAL VOC07+12和自制数据集上mAP值分别提高2.84个百分点和3.06个百分点，FPS保持在87.9，对遮挡目标的检测能力显著提升，满足移动端对狭小复杂场景实时检测的需求。

关键词: 狭小空间, 遮挡目标检测, YOLOv4-tiny, 空间注意力, 多尺度特征融合, 自适应非极大抑制

Abstract: Aiming at the problems of a large number of missed detections and classification errors in the light detection network due to the mutual occlusion of objects in narrow spaces, an adaptive non-maximum suppression（A-NMS） multiscale detection method based on YOLOv4-tiny network is proposed. A large-scale feature map optimization approach and a pyramid pooling model are incorporated into the backbone network to enhance the major regional features of hidden objects. To improve the fusing of shallow detail data with high-level semantic information, a two-way pyramidal feature fusion network with embedded spatial attention is designed. A dynamic NMS threshold setting method that correlates the regional object density with the distance factor of the center of the bounding box is proposed and replaces the traditional IoU-NMS algorithm in the post-processing stage to further reduce the missed detection. The experimental results show that compared with the YOLOv4-tiny algorithm, the improved algorithm improves the mAP value by 2.84 percentage points and 3.06 percentage points on the public dataset PASCAL VOC07+12 and the self-made dataset, respectively, while the FPS remains at 87.9, the detection capability of occluded objects is significantly improved to meet the demand of mobile terminals for real-time detection of narrow and complex scenes.

Key words: narrow space, occluded object detection, YOLOv4-tiny, spatial attention, multi-scale feature fusion, adaptive non-maximum suppression

王长清, 贺坤宇, 蒋帅. 改进YOLOv4-tiny网络的狭小空间目标检测方法[J]. 计算机工程与应用, 2022, 58(10): 240-248.

WANG Changqing, HE Kunyu, JIANG Shuai. Narrow Space Object Detection Method by Improved YOLOv4-tiny Network[J]. Computer Engineering and Applications, 2022, 58(10): 240-248.

参考文献

[1] 华志超.基于监控视频的电梯轿厢内进入目标检测算法研究与实现[D].南京：东南大学，2019.
HUA Z C.Research and implementation of detection algorithm for entering object in elevator car based on surveillance video[D].Nanjing：Southeast University，2019.
[2] 张开生，郭碧筱，刘泽新，等.基于人流量检测的改进CN算法[J].计算机工程与设计，2020，41（2）：411-416.
ZHANG K S，GUO B X，LIU Z X，et al.Improved CN algorithm based on pedestrain flow detection[J].Computer Engineering and Design，2020，41（2）：411-416.
[3] 孙振.电梯轿厢内异常行为检测及其监控系统设计[D].徐州：中国矿业大学，2020.
SUN Z.Abnormal behavior detection and monitoring system design in elevator car[D].Xuzhou：China University of Mining and Technology，2020.
[4] SAUNDERS C，STITSON M O，WESTON J，et al.Support vector machine[J].Computer Science，2002，1（4）：1-28.
[5] 张媛，臧坤，华志超，等.基于电梯监控视频的轿厢中狗识别的算法研究[J].机械设计与制造工程，2018，47（3）：103-107.
ZHANG Y，ZANG K，HUA Z C，et al.Research on algorithm of dog recognition in elevator based on video[J].Mechanical Design and Manufacturing Engineering，2018，47（3）：103-107.
[6] KRIZHEVSKY A，SUTSKEVER I，HINTON G E.ImageNet classification with deep convolutional neural networks[J].Communications of the ACM，2017，60（6）：84-90.
[7] LIU W，ANGUELOY D，ERHAN D，et al.SSD：Single shot multibox detector[C]//14th European Conference on Computer Vision，2016：21-37.
[8] 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，Las Vegas，Jun 27-30，2016：779-788.
[9] REDMON J，FARHADI A.YOLOv3：an incremental improvement[C]//2018 IEEE Conference on Computer Vision and Pattern Recognition，2018：1-4.
[10] 王国新，王珂硕.基于注意力机制的煤矿井下行人检测的轻量化网络结构[J].黑龙江科技大学学报，2021，31（6）：824-829.
WANG G X，WANG K S.Lightweight network structure of underground pedestrian detection in coal mine based on attention mechanism[J].Journal of Heilongjiang University of Science ＆ Technology，2021，31（6）：824-829.
[11] 李海滨，孙远，张文明，等.基于YOLOv4-tiny的溜筒卸料煤尘检测方法[J].光电工程，2021，48（6）：73-86.
LI H B，SUN Y，ZHANG W M，et al.The detection method for coal dust caused by chute discharge based on YOLOv4-tiny[J].Opto-Electronic Engineering，2021，48（6）：73-86.
[12] 华志超，章国宝.基于YOLO的电梯轿厢中狗识别算法[J].工业控制计算机，2019，32（6）：15-16.
HUA Z C，ZHANG G B.YOLO-based algorithm of dog recognition in elevator[J].Industrial Control Computer，2019，32（6）：15-16.
[13] 王琳，卫晨，李伟山，等.结合金字塔池化模块的YOLOv2的井下行人检测[J].计算机工程与应用，2019，55（3）：133-139.
WANG L，WEI C，LI W S，et al.Pedestrian detection based on YOLOv2 with pyramid pooling module in underground coal mine[J].Computer Engineering and Applications，2019，55（3）：133-139.
[14] LIN T Y，DOLLAR P，GIRSHICK R，et al.Feature pyramid networks for object detection[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition，2017：936-944.
[15] BOCHKOVSKIY A，WANG C Y，LIAO H Y M.YOLOv4：optimal speed and accuracy of object detection[C]//2020 IEEE Conference on Computer Vision and Pattern Recognition，2020.
[16] 阳珊，王建，胡莉，等.改进RetinaNet的遮挡目标检测算法研究[J/OL].计算机工程与应用（2021-11-15）[2021-12-27].http：//kns.cnki.net/kcms/detail/11.2127.TP.20211112.
1633.012.html.
YANG S，WANG J，HU L，et al.Research on improved RetinaNet’s occluded object detection algorithm[J/OL].Computer Engineering and Applications（2021-11-15）[2021-12-27].http：//kns.cnki.net/kcms/detail/11.2127.TP.20211112.
1633.012.html.
[17] WANG C Y，LIAO H Y M，WU Y H，et al.CSPNet：a new backbone that can enhance learning capability of CNN[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition workshops，Seattle，Jun 14-19，2020：1571-1580.
[18] ZHAO H，SHI J，QI X，et al.Pyramid scene parsing network[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition，2017：6230-6239.
[19] HOWARD A G，ZHU M，CHEN B，et al.MobileNets：efficient convolutional neural networks for mobile vision applications[J].arXiv：1704.04861，2017.
[20] HOU Q，ZHOU D，FENG J.Coordinate attention for efficient mobile network design[J].arXiv：2103.02907v1，2021.
[21] 王兵，乐红霞，李文璟，等.改进YOLO轻量化网络的口罩检测算法[J].计算机工程与用，2021，57（8）：62-69.
WANG B，LE H X，LI W J，et al.Mask detection algorithm based on improved YOLO lightweight network[J].Computer Engineering and Applications，2021，57（8）：62-69.