YOLO v4框架下Multi-Patch多帧增量式交通视频目标检测

doi:10.13474/j.cnki.11-2246.2022.0138

测绘通报 ›› 2022, Vol. 0 ›› Issue (5): 38-44.doi: 10.13474/j.cnki.11-2246.2022.0138

YOLO v4框架下Multi-Patch多帧增量式交通视频目标检测

文奴^1,2,3, 郭仁忠^1,2,3, 贺彪^1,2,3,4, 万远⁵

1. 深圳大学建筑与城市规划学院, 广东深圳 518061;
2. 深圳大学智慧城市研究院, 广东深圳 518061;
3. 粤港澳智慧城市联合实验室, 广东深圳 518061;
4. 城市国土资源监测与仿真重点实验室, 广东深圳 518034;
5. 湖北师范大学城市与环境学院, 湖北黄石 435002

收稿日期:2021-06-07 修回日期:2022-02-25 发布日期:2022-06-08
通讯作者: 贺彪。E-mail:whu_hebiao@hotmail.com
作者简介:文奴(1989-),男,博士,从事交通视频目标检测、图像处理及计算机视觉相关研究。E-mail:wennu1989@126.com
基金资助:
广东省科技创新战略专项(2020B1212030009);自然资源部城市国土资源监测与仿真重点实验室开放基金(KF-2018-03-031)

Multi-Patch multi-frame incremental traffic video object detection method based on YOLO v4

WEN Nu^1,2,3, GUO Renzhong^1,2,3, HE Biao^1,2,3,4, WAN Yuan⁵

1. School of Architecture & Urban Planning, Shenzhen University, Shenzhen 518061, China;
2. Research Institute for Smart Cities, Shengzhen University, Shenzhen 518061, China;
3. Guangdong-Hong Kong-Macau Joint Laboratory for Smart Cities, Shenzhen 518061, China;
4. Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources, Shenzhen 518034, China;
5. College of Urban and Environmental Sciences, Hubei Normal University, Huangshi 435002, China

Received:2021-06-07 Revised:2022-02-25 Published:2022-06-08

摘要/Abstract

摘要： 提升目标检测模型的泛化能力是计算机视觉领域的研究热点和关键难点。本文提出了一种Multi-Patch方法和多帧增量式预测策略,提升了不同场景下交通视频目标检测的稳健性,有效解决了目标尺度多变导致的视频中目标召回率低的问题。根据视频分辨率和目标尺寸,基于Multi-Patch方法自动将视频帧分割成最佳输入尺寸,使用YOLO v4神经网络并关联连续帧的上下文信息,采用增量式预测策略降低视频目标检测的漏检率,提升不同场景下视频目标的检测置信度得分和召回率。采集不同拍摄条件下的交通视频,验证该方法的有效性。试验结果表明,本文提出的目标检测方法召回率在80%以上,置信度平均得分在0.84以上。

关键词: 视频目标检测, 多帧融合, YOLO v4, 卷积神经网络

Abstract: Improving the generalization ability of object detection model is a research focus and key issue in the field of computer vision. This paper proposes a Multi-Patch method and a multi-frame incremental prediction strategy to improve the robustness of traffic video object detection in different scenarios, and effectively solve the problem of low object recall ratio in videos caused by variable object scales. According to the video resolution and object size, the video frame is automatically divided into the best input size based on the Multi-Patch method, the YOLO v4 neural network is used to correlate the context information of the continuous frame, and the incremental prediction strategy is used to reduce the missed detection rate of the video object detection, and to improve the detection confidence score and recall rate of video object in different scenarios. Collect traffic videos under different shooting conditions to verify the effectiveness of the algorithm. Experimental results show that the object detection method proposed in this paper has a recall rate of more than 80% and an average confidence score of more than 0.84.

Key words: video object detection, multi-frame fusion, YOLO v4, convolutional neural networks

中图分类号:

P237

文奴, 郭仁忠, 贺彪, 万远. YOLO v4框架下Multi-Patch多帧增量式交通视频目标检测[J]. 测绘通报, 2022, 0(5): 38-44.

WEN Nu, GUO Renzhong, HE Biao, WAN Yuan. Multi-Patch multi-frame incremental traffic video object detection method based on YOLO v4[J]. Bulletin of Surveying and Mapping, 2022, 0(5): 38-44.

参考文献

[1] REN S Q, HE K M, 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.
[2] HE K M, GKIOXARI G, DOLLÁR P, et al. Mask R-CNN[C]//Proceedings of 2017 IEEE International Conference on Computer Vision. Venice:IEEE, 2017:2980-2988.
[3] LIU W, ANGUELOV D, ERHAN D, et al. SSD:single shot multibox detector[C]//Proceedings of 2016 European Conference on Computer Vision. Cham:Springer, 2016:21-37.
[4] REDMON J, FARHADI A. YOLO v3:an incremental improvement[EB/OL].[2021-06-07]. http://alumni.soe.ucsc.edu/~czczycz/src/YOLO v3.pdf.
[5] YOO D, PARK S, LEE J Y, et al. Attentionnet:aggregating weak directions for accurate object detection[C]//Proceedings of 2015 IEEE International Conference on Computer Vision. Santiago:IEEE, 2015:2659-2667.
[6] CAICEDO J C, LAZEBNIK S. Active object localization with deep reinforcement learning[C]//Proceedings of 2015 IEEE International Conference on Computer Vision. Santiago:IEEE, 2015:2488-2496.
[7] BOCHKOVSKIY A,WANG C Y,LIAO H Y M.YOLO v4:optimal speed and accuracy of object detection[EB/OL].[2021-06-07]. https://arxiv.org/pdf/2004.10934.pdf.
[8] HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas:2016:770-778.
[9] HE K M, ZHANG X Y, REN S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9):1904-1916.
[10] LIU S T, HUANG D, WANG Y H. Learning spatial fusion for single-shot object detection[EB/OL].[2021-06-07]. https://arxiv.org/pdf/1911.09516.pdf.
[11] ZHOU Y, JI J, SONG K. A moving target detection method based on improved frame difference background modeling[J]. Open Cybernetics&Systemics Journal, 2014, 8(1):970-975.
[12] 乐英,赵志成.基于背景差分法的多运动目标检测与分割[J].中国工程机械学报, 2020, 18(4):305-309.
[13] 欧阳玉梅.基于稠密光流算法的运动目标检测的Python实现[J].现代电子技术, 2021, 44(1):78-82.
[14] FU C Y,LIU W,RANGA A,et al.DSSD:deconvolutional single shot detector[EB/OL].[2021-06-07]. https://arxiv.org/pdf/1701.06659.pdf.
[15] CHEN Y, ZHANG P, LI Z, et al. Stitcher:feedback-driven data provider for object detectio[EB/OL].[2021-06-07]. https://arxiv.org/pdf/2004.12432v1.pdf.
[16] JIANG J, BAO S, SHI W, et al. Improved traffic sign recognition algorithm based on YOLO v3 algorithm[J]. Journal of Computer Applications, 2020, 40(8):2472-2478.
[17] 金立生,郭柏苍,王芳荣,等.基于改进YOLO v3的车辆前方动态多目标检测算法[J].吉林大学学报(工学版), 2021, 51(4):1427-1436.
[18] 赵媛媛,朱军,谢亚坤,等.改进YOLO v3的视频图像火焰实时检测算法[J].武汉大学学报(信息科学版), 2021, 46(3):326-334.
[19] HATAB M,MALEKMOHAMADI H,AMIRA A.Surface defect detection using YOLO network[C]//Proceedings of 2020 SAI Intelligent Systems Conference. Cham:Springer, 2020:505-515.
[20] LI H C, XIONG P F, AN J, et al. Pyramid attention network for semantic segmentation[EB/OL].[2021-06-07]. https://arxiv.org/pdf/1805.10180.pdf.
[21] WANG C Y, LIAO H Y M, WU Y H, et al. CSPNet:a new backbone that can enhance learning capability of CNN[C]//Proceedings of 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Seattle:IEEE, 2020:390-391.
[22] NEUBECK A, VAN GOOL L. Efficient non-maximum suppression[C]//Proceedings of the 18th International Conference on Pattern Recognition. Hong Kong:IEEE, 2006:850-855.

YOLO v4框架下Multi-Patch多帧增量式交通视频目标检测

Multi-Patch multi-frame incremental traffic video object detection method based on YOLO v4

可视化

摘要/Abstract

引用本文

使用本文

参考文献

相关文章 15

编辑推荐

Metrics

本文评价

[1]	柳林, 孙毅, 李万武. 基于CNN海上钻井平台检测模型的构建及训练算法分析[J]. 测绘通报, 2022, 0(7): 26-32,99.
[2]	蔡柔丹. 一种基于用户异步轨迹的身份识别智能方法[J]. 测绘通报, 2022, 0(7): 158-162,167.
[3]	滕文鑫, 张建辰, 邵杰. 集成时空邻近与卷积网络车道级高精度定位算法[J]. 测绘通报, 2022, 0(6): 1-5,61.
[4]	王胜利, 朱寿红, 蒋毅. 基于CNN模型迁移的OLI影像光伏电池板场景识别[J]. 测绘通报, 2022, 0(2): 5-9.
[5]	耿万轩, 周维勋, 金双根. 基于LDCNN特征提取的多核SVM高分辨率遥感影像场景分类[J]. 测绘通报, 2021, 0(8): 14-21.
[6]	赵恒谦, 杨屹峰, 刘泽龙, 宋瑞. 农作物叶片病害迁移学习分步识别方法[J]. 测绘通报, 2021, 0(7): 34-38.
[7]	贾文翰, 刘越岩. 基于卷积神经网络的高分遥感影像内陆盐沼湿地信息提取[J]. 测绘通报, 2021, 0(6): 89-92.
[8]	段雅鸣, 张锦水, 朱爽. 基于深度卷积神经网络的云检测方法[J]. 测绘通报, 2021, 0(4): 33-39.
[9]	马林飞, 倪欢, 周子涵. 基于注意力机制的高分遥感图像采石场识别[J]. 测绘通报, 2021, 0(11): 96-100.
[10]	彭述刚, 王大成, 谭军辉, 敖旋峰, 俞佳颖. 基于改进卷积神经网络的城市地下排水管道视频智能识别[J]. 测绘通报, 2021, 0(10): 132-135,167.
[11]	王大成, 谭军辉, 彭述刚, 钟镇声, 陈国强, 李国桥. 利用深度学习模型智能识别地下排水管道缺陷[J]. 测绘通报, 2021, 0(10): 141-145.
[12]	杨泽宇, 张洪艳, 明金, 冷伟, 刘海启, 游炯. 深度学习在高分辨率遥感影像冬油菜提取中的应用[J]. 测绘通报, 2020, 0(9): 110-113.
[13]	马超. 融合卷积神经网络和循环神经网络的车轮目标检测[J]. 测绘通报, 2020, 0(8): 139-143.
[14]	何肖娜, 宋斌斌, 余敏. 融合上下文感知的地标检测辅助WiFi-PDR室内定位[J]. 测绘通报, 2020, 0(6): 7-11,16.
[15]	赵琪, 孙立双, 谢志伟. 基于改进MCNN的密度图在室内定位中的应用[J]. 测绘通报, 2020, 0(6): 12-16.