摘要
为提高恶劣雨天交通环境下车辆目标检测精度,提出一种基于PReNet和YOLOv4融合的深度学习网络DTOD-PReYOLOv4,融合了改进的图像复原子网D-PReNet和改进的目标检测子网TOD-YOLOv4;将多尺度膨胀卷积融合模块和添加了挤压激励块的注意机制残差模块引入PReNet,获得的D-PReNet可更有效提取雨纹特征;使用轻量化的CSPDarknet26代替YOLOv4骨干模块CSPDarknet53,为YOLOv4的颈部路径聚合网络模块添加复合残差块,同时采用k-means++算法取代原始网络聚类算法,获得的TOD-YOLOv4可在改善交通小目标检测精度的同时进一步提高检测效率;基于构建的雨天交通场景车辆目标数据集VOD-RTE对DTOD-PReYOLOv4进行了验证。研究结果表明:与当前YOLO系列主流网络相比,提出的DTOD-PReYOLOv4对原浅层ResBlock_body1叠加残差块,可以更好地提取分辨率较小的特征;对原深层ResBlock_body3、ResBlock_body4和ResBlock_body5进行裁剪,获得ResBlock_body3×2、ResBlock_body4×2和ResBlock_body5×2,可以有效降低卷积层冗余,提高内存利用率;为PANet中的Concat+Conv×5添加跳跃连接形成CRB模块,可以有效缓解网络层数加深引起的小目标检测效果退化问题;采用k-means++算法,在多尺度检测过程中为较大的特征图分配更加适合的较小先验框,为较小的特征图分配更加适合的较大先验框,进一步提高了目标检测的精度;与MYOLOv4相比,精确率和召回率的调和平均值、平均精度、检测速度分别提升了5.02%、6.70%、15.63帧·s^(-1),与TOD-YOLOv4相比,分别提升了3.51%、4.31%、2.17帧·s^(-1),与YOLOv3相比,分别提升了46.07%、48.05%、18.97帧·s^(-1),与YOLOv4相比,分别提升了31.06%、29.74%、16.26帧·s^(-1)。
In order to improve the detection accuracy of vehicle target in severe rainy day under traffic environment, a deep learning network DTOD-PReYOLOv4(derain and traffic object detection-PReNet and YOLOv4) was proposed based on the fusion of PReNet and YOLOv4, which integrated the improved image restoration subnet D-PReNet and the improved target detection subnet TOD-YOLOv4. D-PReNet could extract rain streak features more effectively, since it introduced the multi-scale expansion convolution fusion module(MSECFM) and the attentional mechanism residual module(AMRM) with SEBlock into PReNet. TOD-YOLOv4 improved not only the detection accuracy of small traffic target, but also the detection efficiency, since it replaced the backbone module CSPDarknet53 of YOLOv4 with the lightweight CSPDarknet26 of YOLOv4, added CRB into PANet of YOLOv4 neck, and utilized k-means++ instead of the original network clustering algorithm. DTOD-PReYOLOv4 was verified based on the constructed vehicle target database VOD-RTE in rainy day traffic scenario. Research results show that compared with the current series of YOLO networks, the proposed DTOD-PReYOLOv4 can better extract the features with lower resolutions by superimposing RB over ResBlock_body1 in the shallow layer. It can effectively reduce the convolutional layer redundancy and improve the memory utilization, since ResBlock_body3, ResBlock_body4 and ResBlock_body5 in deep layer can be properly cropped to ResBlock_body3×2, ResBlock_body4×2 and ResBlock_body5×2, respectively. It also can alleviate the degradation of small target detection effect caused by the deepening of network layers by adding jump connection to Concat+Conv×5 in PANet to form CRB. In the process of multi-scale detection, k-means++ algorithm is adopted to allocate smaller prior boxes that are more suitable for the larger feature images, but larger prior boxes that are more suitable for smaller feature images, which further improves the accuracy of target detection. The harmonic mean value of precision and recall rate, average precision and detection speed of DTOD-PReYOLOv4 respectively increase by 5.02%, 6.70% and 15.63 frames per second compared with MYOLOv4, by 3.51%, 4.31% and 2.17 frames per second compared with TOD-YOLOv4, by 46.07%, 48.05% and 18.97 frames per second compared with YOLOv3, and by 31.06%, 29.74% and 16.26 frames per second compared with YOLOv4. 4 tabs, 12 figs, 44 refs.
作者
陈婷
姚大春
高涛
仇会会
郭昶鑫
刘占文
李永会
边浩毅
CHEN Ting;YAO Da-chun;GAO Tao;QIU Hui-hui;GUO Chang-xin;LIU Zhan-wen;LI Yong-hui;BIAN Hao-yi(School of Information Engineering,Chang'an University,Xi'an 710064,Shaanxi,China;Branch of Shaanxi,Bank of Communications Co.,Ltd.,Xi'an 710004,Shaanxi,China;School of Electrical and Information Engineering,The University of Sydney,Sydney NSW2006,New South Wales,Australia;Zhejiang Institute of Mechanical and Electrical Technology,Hangzhou 310053,Zhejiang,China)
出处
《交通运输工程学报》
EI
CSCD
北大核心
2022年第3期225-237,共13页
Journal of Traffic and Transportation Engineering
基金
国家重点研发计划(2019YFE0108300,2018YFB1600600)
国家自然科学基金项目(62001058,52172379)
中央高校基本科研业务费专项资金项目(300102241201,300102242901)
浙江省科技厅软科学项目(2021C25005)
浙江省交通运输厅科技计划项目(2021032)。
