基于Faster R-CNN的无人机车辆目标检测

doi:10.16180/j.cnki.issn1007-7820.2021.11.002

摘要/Abstract

摘要：

无人机视角目标存在分辨率低、完整度低、干扰项多等缺点。此外,无人机目标检测系统研究进展缓慢,其对于小目标、不完整目标和重叠目标的检测精度无法满足社会实际需求。针对以上问题,文中提出一种基于Faster R-CNN的无人机平台车辆目标检测解决方案。该方案使用ResNet卷积神经网络作为特征提取网络,并改进网络结构,重新设计Anchor生成和改进Soft-NMS算法等策略,解决了小目标、不完整目标和重叠目标的检测精度低等问题, 提高了无人机车辆检测的精度。文中所构建的数据集测试实验表明,所提算法较改进前AP值提高13.46%。公开数据集上的测试实验表明,相较于目前的主流算法,文中所提算法拥有更好的AP值和召回率。

关键词: Faster R-CNN, 无人机图像, 车辆检测, ResNet, 卷积神经网络, 网络结构改进, Anchor生成, Soft-NMS算法

Abstract:

There are disadvantages such as low resolution, low completeness, and many interference items in the UAV's perspective targets. Additionally, the research progress of UAV target detection system is slow, and its detection accuracy for small targets, incomplete targets and overlapping targets cannot meet the actual needs of society. In view of these problems, this study proposes a vehicle target detection solution for UAV platform based on Faster R-CNN. This solution uses ResNet convolutional neural network as the feature extraction network, improves the network structure, redesigns Anchor generation and improves the Soft-NMS algorithm and other strategies, solves the problem of low detection accuracy of small targets, incomplete targets and overlapping targets, and improves the accuracy of UAV vehicle detection. The test experiments on the dataset constructed in this study show that the proposed algorithm has 13.46% increase in AP value compared with the previous improvement. Test experiments on the public data set show that compared with the current mainstream algorithms, the proposed algorithm has better AP value and recall rate.

Key words: Faster R-CNN, UAV image, vehicle detection, ResNet, convolutional neural network, network structure improvement, Anchor generation, Soft-NMS algorithm

中图分类号:

TP391

张莹,刘子龙,万伟. 基于Faster R-CNN的无人机车辆目标检测[J]. 电子科技, 2021, 34(11): 11-20.

ZHANG Ying,LIU Zilong,WAN Wei. UAV Vehicle Target Detection Based on Faster R-CNN[J]. Electronic Science and Technology, 2021, 34(11): 11-20.

图/表 26

图1

图2

图3

表1

图4

图5

表2

图6

图7

图8

表3

图9

图10

图11

图12

图13

图14

表4

表5

表6

表7

图15

图16

图17

图18

表8

参考文献 20

[1]	袁小平, 王岗, 王晔枫, 等. 基于改进卷积神经网络的交通标志识别方法[J]. 电子科技, 2019, 32(11):28-32.
	Yuan Xiaoping, Wang Gang, Wang Yefeng, et al. Traffic sign recognition method based on improved convolutional neural network[J]. Electronic Science and Technology, 2019, 32(11):28-32.
[2]	Joseph R, Santosh D, Ross G, et al. You only look once: unified, real-time object detection[C]. Las Vegas:IEEE Conference on Computer Vision and Pattern Recognition, 2016.
[3]	Joseph R, Ali F. YOLO9000:Better,faster,stronger[C]. Honolulu:IEEE Conference on Computer Vision and Pattern Recognition, 2016.
[4]	Jiao Z T, Zhang Y M, Mu L X, et al. A YOLOv3-based learning strategy for real-time UAV-based forest fire detection[C]. Hefei:The Thirty-second Conference on Control and Decision-Making in China, 2020.
[5]	Wei L, Dragomir A, Dumitru E, et al. SSD:Single shot multibox detector[C]. London:European Conference on Computer Vision, 2016.
[6]	Ross G, Jeff D, Trevor D, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]. Columbus:IEEE Conference on Computer Vision and Pattern Recognition, 2014.
[7]	Ross G. Fast R-CNN[C]. Santiago:IEEE International Conference on Computer Vision, 2015.
[8]	Shaoqing R, Kaiming H, Ross G, 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. doi: 10.1109/TPAMI.2016.2577031 pmid: 27295650
[9]	Zhu P, Wen L, Du D, et al. Visdron-det2018:The vision meets drone object detetion in image challenge result[C]. Munich:Proceedings of the European Conference on Computer Vision, 2018.
[10]	Mundhenk T N, Konjevod G, Sakla W A, et al. A large contextual dataset for classification, detection and counting of cars with deep learning[C]. Amsterdam:European Conference on Computer Vision, 2016.
[11]	Sébastien R, Frédéric J. Vehicle detection in aerial imagery:a small target detection benchmark[J]. Journal of Visual Communication and Image Representation, 2016, 34(7):187-203. doi: 10.1016/j.jvcir.2015.11.002
[12]	Yang M Y, Liao W, Li X, et al. Vehicle detection in aerial images[J]. American Society for Photogrammetry and Remote Sensing, 2019, 85(4):297-304.
[13]	Zhu H G, Chen X G, Dai W Q, et al. Orientation robust object detection in aerial images using deep convolutional neural network[C]. Quebec City:IEEE International Conference on Image Processing, 2015.
[14]	De A, Paulo R L O, Luiz S B, et al. PKLot-A robust dataset for parking lot classification[J]. Expert Systems with Applications, 2015, 42(11):4937-4949. doi: 10.1016/j.eswa.2015.02.009
[15]	周文凯, 韩芳, 孔维健. 基于Faster-RCNN的极验点选式验证码识别[J]. 电子科技, 2019, 32(9):42-46.
	Zhou Wenkai, Han Fang, Kong Weijian. Point-selective geetest captcha recognition based on Faster-RCNN[J]. Electronic Science and Technology, 2019, 32(9):42-46.
[16]	He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]. Las Vegas:IEEE Conference on Computer Vision and Pattern Recognition, 2016.
[17]	Alexander N, Luc V G. Efficient non-maximum suppression[C]. Hong Kong:The Eighteenth International Conference on Pattern Recognition, 2006.
[18]	赵文清, 严海, 邵绪强. 改进的非极大值抑制算法的目标检测[J]. 中国图象图形学报, 2018(11):1676-1685.
	Zhao Wenqing, Yan Hai, Shao Xuqiang. Target detection based on improved non maximum suppression algorithm[J]. Journal of Image and Graphics, 2018(11):1676-1685.
[19]	Navaneeth B, Bharat S, Rama C, et al. Soft-NMS--Improving object detection with one line of code[C]. Venice:IEEE International Conference on Computer Vision, 2017.
[20]	杨超, 周大可, 杨欣. 基于检测-分割的图像拼接篡改盲取证算法[J]. 电子设计工程, 2020, 28(13):169-174.
	Yang Chao, Zhou Dake, Yang Xin. Spliced image blind forensics based on detection and segmentation[J]. Electronic Design Engineering, 2020, 28(13):169-174.

数据集	采集设备	分辨率	多场景	车辆图像数目	车辆目标数目
COWC^[10]	航空器	低	是	-	32 716
VEDAI^[11]	卫星	低	是	982	2 950
ITCVD^[12]	飞机	高	是	173	29 088
UCAS-AOD^[13]	卫星	低	是	510	7 114
PUCCR^[14]	摄像头	高	否	2 315	192 216
本文数据集	无人机	高	是	19 522	377 841

Anchor数量	Anchor组合	AP值/%
3	{16²,32²,64²}	91.81
3	{32²,64²,96²}	90.80
3	{64²,96²,128²}	89.19
4	{16²,32²,64²,96²}	93.54
4	{32²,64²,96²,128²}	91.97
5	{16²,32²,64²,96²,128²}	94.03

	α=0.3	α=0.4	α=0.5	α=0.6	α=0.7	α=0.8	α=0.9	α=1.0
改进Soft-NMS算法1	91.57	91.44	91.26	91.81	91.32	91.27	91.31	91.08
改进Soft-NMS算法2	91.32	91.03	91.09	91.47	91.48	91.36	91.01	91.77

算法	AP值/%	召回率/%
Faster R-CNN (VGG16)	74.68	75.58
YOLOv3 (darknet53)	72.55	72.98
SSD (VGG16)	67.69	70.76
R-FCN (ResNet101)	79.42	83.21
本文算法	82.94	86.23