工程科学学报.第44卷，第X期：1-11.2021年X月 Chinese Journal of Engineering,Vol.44,No.X:1-11,X 2021 https://doi.org/10.13374/j.issn2095-9389.2020.10.28.006;http://cje.ustb.edu.cn 基于梯度压缩的YOLO V4算法车型识别 牟亮，赵红四，李燕，仇俊政，孙传龙，刘晓童 青岛大学机电工程学院，青岛266071 ☒通信作者，E-mail:qdlizh@163.com 摘要为进一步提高智能交通系统对车辆及不同车型识别的泛化性、鲁棒性与实时性.根据检测区域的特征有针对性地 构建数据集，改变余弦退火衰减(CD)学习率的更新方式，提出一种基于梯度压缩(GC)的Adam优化算法(Adam-GC)来提高 YOLO V4算法的训练速度、检测精度以及网络模型的泛化能力.为验证改进后YOLO v4算法的有效性，对实际路况的车流 进行采集后，利用训练完成的网络模型对不同密度车流进行定量的车型检测实验验证.经实验验证，改进后方法的整体检测 结果要优于改进前，YOL0v4和YOL0v4GCCD训练得到的网络模型在阻塞流样本下检测得到的准确率分别为94.59%和 96.46%;在同步流样本下检测得到的准确率分别为95.34%和97.20%：在自由流样本下检测得到的准确率分别为95.98%和 97.88%. 关键词梯度压缩：学习率；Adam优化算法：YOLO v4:车型识别 分类号TP391.4 Vehicle recognition based on gradient compression and YOLO v4 algorithm MU Liang,ZHAO Hong,LI Yan,OIU Jun-zheng,SUN Chuan-long,LIU Xiao-Tong College of Mechanical and Electrical Engineering,Qingdao University,Qingdao 266071,China Corresponding author,E-mail:qdlizh@163.com ABSTRACT Intelligent transportation systems (ITS)are the development direction of future transportation systems.ITS can effectively reduce traffic load and environmental pollution and ensure traffic safety,which has been a concern in all countries.In the field of intelligent transportation,vehicle detection has always been a hot spot but a difficult matter.To further improve the generalization,robustness,and real-time performance of the intelligent transportation system for the recognition of vehicles and different vehicle types,this study proposes an improved vehicle detection algorithm and chooses a road in the city as the background of the article. According to the characteristics of the detection region,the data set is constructed pertinently and the data set size is reduced using a video frame extraction method,aiming at achieving better detection performance with less training cost.The updating method of cosine decay with warm-up(CD)learning rate is then changed.An Adam gradient compression(GC)based on GC is proposed to improve the training speed,detection accuracy,and generalization ability of the YOLO v4 algorithm.To verify the effectiveness of the proposed algorithm,the trained network model is used to verify the quantitative vehicle type detection experiment of different density traffic flows after collecting the traffic flow information under actual road conditions.Experimental results show that the overall detection of the improved method is better than that of the original method.The accuracy rates of the network models trained by YOLO v4 and YOLO v4 GC CD under the blocking flow samples,synchronous flow samples,and free flow samples are 94.59%and 96.46%,95.34%and 97.20%,95.98%,and 97.88%,respectively.Simultaneously,the detection effect of YOLOV4 GC CD was verified at night and on rainy days with an accuracy rate of 92.06%and 95.51%,respectively. KEY WORDS gradient compression;learning rate;Adam optimization algorithm;YOLO v4;vehicle recognition 收稿日期：2020-10-28 基金项目：青岛市民生科技计划资助项目(19-6-1-88-nsh),山东省重点研发计划资助项目(2018GGX105004)基于梯度压缩的 YOLO v4 算法车型识别 牟 亮，赵 红苣，李 燕，仇俊政，孙传龙，刘晓童 青岛大学机电工程学院, 青岛 266071 苣通信作者， E-mail: qdlizh@163.com 摘    要    为进一步提高智能交通系统对车辆及不同车型识别的泛化性、鲁棒性与实时性. 根据检测区域的特征有针对性地 构建数据集，改变余弦退火衰减（CD）学习率的更新方式，提出一种基于梯度压缩（GC）的 Adam 优化算法（Adam−GC）来提高 YOLO v4 算法的训练速度、检测精度以及网络模型的泛化能力. 为验证改进后 YOLO v4 算法的有效性，对实际路况的车流 进行采集后，利用训练完成的网络模型对不同密度车流进行定量的车型检测实验验证. 经实验验证，改进后方法的整体检测 结果要优于改进前，YOLO v4 和 YOLO v4 GC CD 训练得到的网络模型在阻塞流样本下检测得到的准确率分别为 94.59% 和 96.46%；在同步流样本下检测得到的准确率分别为 95.34% 和 97.20%；在自由流样本下检测得到的准确率分别为 95.98% 和 97.88%. 关键词    梯度压缩；学习率；Adam 优化算法；YOLO v4；车型识别 分类号    TP391.4 Vehicle recognition based on gradient compression and YOLO v4 algorithm MU Liang，ZHAO Hong苣 ，LI Yan，QIU Jun-zheng，SUN Chuan-long，LIU Xiao-Tong College of Mechanical and Electrical Engineering, Qingdao University, Qingdao 266071, China 苣 Corresponding author, E-mail: qdlizh@163.com ABSTRACT    Intelligent transportation systems (ITS) are the development direction of future transportation systems. ITS can effectively reduce traffic load and environmental pollution and ensure traffic safety, which has been a concern in all countries. In the field of intelligent transportation, vehicle detection has always been a hot spot but a difficult matter. To further improve the generalization, robustness, and real-time performance of the intelligent transportation system for the recognition of vehicles and different vehicle types, this study proposes an improved vehicle detection algorithm and chooses a road in the city as the background of the article. According to the characteristics of the detection region, the data set is constructed pertinently and the data set size is reduced using a video frame extraction method, aiming at achieving better detection performance with less training cost. The updating method of cosine decay with warm-up (CD) learning rate is then changed. An Adam gradient compression (GC) based on GC is proposed to improve the training speed, detection accuracy, and generalization ability of the YOLO v4 algorithm. To verify the effectiveness of the proposed algorithm, the trained network model is used to verify the quantitative vehicle type detection experiment of different density traffic flows after collecting the traffic flow information under actual road conditions. Experimental results show that the overall detection of the improved method is better than that of the original method. The accuracy rates of the network models trained by YOLO v4 and YOLO v4 GC CD under the blocking flow samples, synchronous flow samples, and free flow samples are 94.59% and 96.46%, 95.34% and 97.20%, 95.98%, and 97.88%, respectively. Simultaneously, the detection effect of YOLOV4 GC CD was verified at night and on rainy days with an accuracy rate of 92.06% and 95.51%, respectively. KEY WORDS    gradient compression；learning rate；Adam optimization algorithm；YOLO v4；vehicle recognition 收稿日期: 2020−10−28 基金项目: 青岛市民生科技计划资助项目（19-6-1-88-nsh），山东省重点研发计划资助项目（2018GGX105004） 工程科学学报，第 44 卷，第 X 期：1−11，2021 年 X 月 Chinese Journal of Engineering, Vol. 44, No. X: 1−11, X 2021 https://doi.org/10.13374/j.issn2095-9389.2020.10.28.006; http://cje.ustb.edu.cn