\r高春艳，赵文辉，张明路，孟宪春\r
\r
AuthorsHTML:\r高春艳，赵文辉，张明路，孟宪春\r
\r
AuthorsListE:\rGao Chunyan，Zhao Wenhui，Zhang Minglu，Meng Xianchun\r
\r
AuthorsHTMLE:\rGao Chunyan，Zhao Wenhui，Zhang Minglu，Meng Xianchun\r
\r
Unit:\r河北工业大学机械工程学院，天津 300130\r
\r
Unit_EngLish:\rSchool of Mechanical Engineering，Hebei University of Technology，Tianjin 300130，China\r
\r
Abstract_Chinese:\r\r在公共安防领域，汽车底部潜藏的危险品危害性强，检测难度大．当前车底危险品检测主要通过模板匹配等传统目标检测技术进行检测，但存在检测速度慢、检测精度低的问题，为了能够更好地检测出藏匿于车底部位的危险品目标，提出一种改进的\rYOLOv3\r目标检测算法．该方法分别从多尺度图像训练、增加\rInception\r-\rres\r模块和省去大尺度特征输出分支\r3\r个方面对\rYOLOv3\r网络进行改进．实验证明：在自制危险品数据集下，采用双数据集多尺度图像训练，网络的\rmAP\r值大约提高了\r0.9\r%\r，单张图像检测耗时大致不变；在\r3\r个支路分别增加相应\rInception\r-\rres\r结构，网络的\rmAP\r值大约提高了\r1.5\r%\r，但是单张图像检测耗时却增加了原来的\r2.6\r倍；省去大尺度特征输出分支，网络的\rmAP\r值降低了\r0.3\r%\r，但是单张图像检测耗时也相应降低\r25.4\r%\r．通过结合上述方法对\rYOLOv3\r算法模型进行综合改进，选取双数据集多尺度图像训练的方式，同时省去大尺度特征输出分支，并在其他两支路增加相应\rInception\r-\rres\r结构．这样在充分结合\rInception\r-\rres\r结构优势的情况下，省去对检测耗时影响较大且对检测结果\rmAP\r值影响较小的大尺度特征输出分支．实验测得改进网络\rmAP\r值大约提高\r2.2\r%\r左右，而单张图像检测耗时增加了\r0.014s\r，在可接受范围内．且网络对于小尺寸目标识别效果明显增强，很好地满足了车底复杂背景危险品检测要求\r.\r\r
\r
Abstract_English:\r\rIn the field of public security\r，\rdangerous objects hidden at the bottom of a vehicle are highly harmful and difficult to detect. In the field of vehicle bottom dangerous object detection\r，\rtraditional object detection technology\r，\rsuch as template matching\r，\ris mainly used\r．\rHowever\r，\rthe detection speed is slow and the detection accuracy is low. To better detect dangerous objects hidden at the bottom of a vehicle\r，\ran improved YOLOv3 detection algorithm is proposed. The method improves three aspects of the YOLOv3 network\r，\ri.e.\r，\rmulti-scale image training\r，\radding the Inception-res module\r，\rand eliminating the large-scale feature output branch. The experiment proves that\r，\runder the self-made dangerous object dataset\r，\rusing the double-dataset multi-scale image training\r，\rthe mAP value of the network increases by approximately 0.9\r%\r，\rbut the detection time of a single image remains roughly the same. When adding the corresponding Inception-res structure to the three branches\r，\rthe network’s mAP value increases by approximately 1.5\r%\r，\rbut the detection time of a single image increases by 2.6 times. When eliminating the large-scale feature output branch\r，\rthe network’s mAP value decreases by 0.3\r%\r，\rbut the detection time of a single image decreases by 25.4\r%\r. By combining the three aspects\r，\ri.e.\r，\radopting the double-dataset multi-scale image training method\r，\reliminating the large-scale feature output branch\r，\rand adding the corresponding Inception-res structure to the two other branches\r，\rthe YOLOv3 algorithm model is comprehensively improved. In this manner\r，\rin combination with the advantages of the Inception-res structure\r，\rthe large-scale feature output branch that has a considerable effect on the detection time and has only a slight influence on the mAP value of the detection result is omitted. The experimental results show that the mAP value of the improved network increases by approximately 2.2\r%\r. Meanwhile\r，\rthe detection time of a single image increases by 0.014s\r，\rwhich is within the acceptable range. Moreover\r，\rthe network has significantly enhanced the recognition effect on small-sized objects\r，\rwhich satisfies the requirements for the detection of dangerous objects in the complex background of a vehicle.\r\r
\r
Keyword_Chinese:深度学习；卷积神经网络；YOLOv3 算法；危险品检测\r

Keywords_English:deep learning；convolutional neural network；YOLOv3 algorithm；dangerous object detection\r


PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6435