一种改进YOLOv3的动态小目标检测方法

doi:10.19665/j.issn1001-2400.2020.03.001

摘要/Abstract

摘要：

由于低空小型无人机类的动态小目标视觉特征不明显,且在检测过程中尺度可能变化较大,故传统的检测算法在检测该类目标时易受到干扰,检测速度和稳定性较差。针对此问题,提出了一种结合YOLOv3改进模型和超分辨率重建技术的无人机实时检测算法。首先以三帧间差分法筛选可疑区域; 然后使用轻量级卷积神经网络进行可疑区域的超分辨率重建,增强细节信息;再用维度聚类算法重新生成YOLOv3模型的预选框参数并调整预选框分配,使用改进模型扫描全图和可疑区域,进行无人机检测;在视频流检测中,将帧间关系作为依据,强化选定区域的细节特征后再进行目标检测,实现无人机的检测式追踪。该方法在GTX1070Ti处理器加速下, 检测速度可达18帧每秒,模型检测的准确率和召回率分别为96.8%和95.6%。实验结果表明,该方法可以在复杂环境下检测大疆精灵系列无人机,检测有效距离可观,相比传统算法和机器学习类特征提取算法,其处理速度和鲁棒性更佳。

关键词: 反无人机, 卷积神经网络, 超分辨率, 复杂背景, 实时检测

Abstract:

The visual characteristics of low-altitude drones are less obvious and the scale changes during the detection process. Traditional detection methods are susceptible to interference during detection, and most of those methods cannot work quickly and robustly. To solve this problem, a real-time drone detection algorithm combined with the improved YOLOv3 model and the super resolution method is proposed in this paper. First, frame difference is used to propose the candidate area, and the super-resolution method is used to strengthen the details. Then the dimensional clustering algorithm is used to regenerate the anchors for the model, and the model is slightly adjusted. Finally, we use the improved YOLOv3 to scan both the whole frame and the processed candidate area so as to detect the drones. The frame relationship is also used to implement tracking of drones by real-time detection. With GPU (GTX 1070Ti) acceleration, the method works at a speed of about 20FPS and has an accuracy rate of 96.8% and a recall rate of 95.6%. The results prove that the method can detect drones in different complex backgrounds with a considerable effective detection distance. Compared with the traditional method or normal machine learning method, our method is of a certain theoretical and practical value.

Key words: anti-drone, convolutional neural network, super resolution, complex background, real-time detection

中图分类号:

TP391.41

崔艳鹏,王元皓,胡建伟. 一种改进YOLOv3的动态小目标检测方法[J]. 西安电子科技大学学报, 2020, 47(3): 1-7.

CUI Yanpeng,WANG Yuanhao,HU Jianwei. Detection method for a dynamic small target using the improved YOLOv3[J]. Journal of Xidian University, 2020, 47(3): 1-7.

图/表 10

图1

图2

图3

图4

图5

图6

表1

表2

表3

图7

参考文献 15

[1]	SHI X, YANG C, XIE W, et al. Anti-drone System with Multiple Surveillance Technologies: Architecture, Implementation, and Challenges[J]. IEEE Communications Magazine, 2018,56(4):68-74.
[2]	HOU X, ZHANG L. Saliency Detection: a Spectral Residual Approach [C]//Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Washington: IEEE Computer Society, 2007: 4270292.
[3]	NISHU S. Motion Detection Based on Frame Difference Method[J]. International Journal of Information and Computation Technology, 2014,4(15):1559-1565.
[4]	官洪运, 汪晨, 苏振涛. 基于ViBe改进的运动目标检测算法[J]. 信息技术与网络安全, 2019(6):36-40.
	GUAN Hongyun, WANG Chen, SU Zhentao. Improved Moving Object Detection Algorithm Based on ViBe[J]. Information Technology and Network Security, 2019(6):36-40.
[5]	王晓. 视频中无人机的实时检测与跟踪算法研究[D]. 哈尔滨: 哈尔滨工程大学, 2019.
[6]	邵盼瑜. 基于视觉的无人机入侵检测与跟踪系统设计与实现[D]. 杭州: 浙江大学, 2018.
[7]	GU J, WANG Z, KUEN J, et al. Recent Advances in Convolutional Neural Networks[J]. Pattern Recognition, 2018,77:354-377. doi: 10.1016/j.patcog.2017.10.013
[8]	REDMON J, FARHADI A. Yolov3: An Incremental Improvement[EB/OL]. [2019-10-31].https://arxiv.org/pdf/1804.02767.pdf.
[9]	DONG C, LOY C C, HE K, et al. Image Super-resolution Using Deep Convolutional Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016,38(2):295-307. doi: 10.1109/TPAMI.2015.2439281
[10]	REDMON J, DIVVALA S, GIRSHICK R , et al. You Only Look Once: Unified, Real-time Object Detection [C]//Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Washington: IEEE Computer Society, 2016: 779-788.
[11]	LIU W, ANGUELOV D, ERHAN D , et al. SSD: Single Shot Multibox Detector [C]//Lecture Notes in Computer Science: 9905. Heidelberg: Springer Verlag, 2016: 21-37.
[12]	REDMON J, FARHADI A. YOLO9000: Better, Faster, Stronger [C]//Proceedings of the 30th IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 6517-6525.
[13]	刘树东, 王晓敏, 张艳. 一种对称残差 CNN 的图像超分辨率重建方法[J]. 西安电子科技大学学报, 2019,46(5):15-23.
	LIU Shudong, WANG Xiaomin, ZHANG Yan. Symmetric Residual Convolution Neural Networks for the Image Super-resolution Reconstruction[J]. Journal of Xidian University, 2019,46(5):15-23.
[14]	LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCo: Common Objects in Context [C]//Lecture Notes in Computer Science: 8693. Heidelberg: Springer Verlag, 2014: 740-755.
[15]	JEONG H J, PARK K S, HA Y G. Image Preprocessing for Efficient Training of YOLO Deep Learning Networks [C]//Proceedings of the 2018 IEEE International Conference on Big Data and Smart Computing. Piscataway: IEEE, 2018: 635-637.

组号	算法	准确率/%	召回率/%	F₁/%	检测速度/(帧·秒^-1)
1	帧间差分法	78.4	79.6	78.9	29.20
2	视觉建模法	83.7	82.5	83.1	16.00
3	HOG+SVM	96.4	93.6	95.0	0.42
4	文中模型	95.8	91.0	93.3	19.00

组数	算法	峰值信噪比/dB	结构相似性/%
1	BICUBIC	24.6	88.0
2	SRCNN	28.0	92.0
3	SRMCNN	28.5	93.3

组号	算法	准确率/%	召回率/%	F₁/%	帧率/(帧·秒)^-1
1	帧差	92.6	90.7	91.6	20.00
2	SVM	98.4	97.8	98.1	5.56
3	文中	96.8	95.6	96.2	18.00