基于图神经网络的车辆目标遮蔽关重部位检测

doi:10.12382/bgxb.2024.0574

摘要/Abstract

摘要：

针对复杂背景和车辆姿态变化导致的车辆关重部位在图像中被遮蔽,无法准确识别的问题,提出了部分可形变物体图卷积神经网络(Partially Deformable Object Graph Convolutional Network,PDO-GCN)的遮蔽车辆关重部位检测模型。该方法以车辆刚体结构关系为基础,构建了基于PDO-GCN的二维成像平面上关重部位之间的空间关联模型,并利用可见关重部位的检测结果估计遮蔽关重部位的位置。实验结果表明,PDO-GCN模型在无需复杂标注的前提下,能够有效推断完整车辆结构信息,显著提高遮蔽部位的检测精度,且满足实时性要求,具有良好的应用潜力。

关键词: 目标检测, 车辆关重部位, 遮蔽, 刚体结构, 先验知识, 图卷积神经网络

Abstract:

The key parts of vehicle occluded due to complex backgrounds and variations in vehicle posture can not be accurately identified in images. A detection method based on partially deformable object graph convolutional network (PDO-GCN) is proposed for detecting the occluded key parts of vehicle. This method is founded on the rigid body structural relationships of vehicles, constructing a spatial association model between key parts on the 2D imaging plane based on PDO-GCN, and utilizes the detected results of visible key parts to estimate the locations of occluded ones. Experimental results demonstrate that the PDO-GCN model can effectively infer the complete vehicle structural information without the need for complex annotations, significantly improves the detection accuracy of occluded parts and fulfils the real-time requirements, thus showcasing considerable potential for practical application.

Key words: object detection, vehicle key part, occlusion, rigid structure, prior knowledge, graph convolutional neural network

中图分类号:

TP301

王烨茹, 杨耿, 刘述, 许啸, 陈华杰, 秦飞巍, 徐华杰. 基于图神经网络的车辆目标遮蔽关重部位检测[J]. 兵工学报, 2024, 45(S1): 242-251.

WANG Yeru, YANG Geng, LIU Shu, XU Xiao, CHEN Huajie, QIN Feiwei, XU Huajie. GCN-based Detection of Occluded Key Parts of Vehicle Target[J]. Acta Armamentarii, 2024, 45(S1): 242-251.

图/表 18

图1 坦克14个关重部位分布

Fig.1 Distribution of 14 key parts of tanks

图2 车辆关重部位检测总体方案

Fig.2 Overall plan for detecting the key parts of vehicles

图3 车辆关重部位对应的GCN网络结构

Fig.3 GCN network structure corresponding to the key parts of vehicle

图4 PDO-GCN结构

Fig.4 PDO-GCN structure

图5 抗混淆特征编码模块

Fig.5 Anti-confusion feature encoding module

图6 基于抗混淆特征编码的图卷积模块

Fig.6 Graph convolution module based on anti-confusion feature encoding

图7 车辆空间坐标示意图

Fig.7 Schematic diagram of vehicle spatial coordinates

图8 场景设置示意图

Fig.8 Scene setting diagram

图9 相机外部参数和车辆形变参数示意图

Fig.9 Schematic diagram of external camera parameters and vehicle deformation parameters

图10 数据集1-无人机拍摄坦克模型示例

Fig.10 Dataset 1-Examples of tank models captured by drones

图11 数据集2-近处坦克数据示例

Fig.11 Dataset 2-Examples of near tank data

表1 消融实验结果(数据集1)

Table 1 Ablation experimental results (Dataset 1)

模块			召回率/%	精确率/%	FPS
YOLOv8	PDO-GCN	Anti Confusion	召回率/%	精确率/%	FPS
√			74.87	81.53	61.3
√	√		13.15	14.23	57.2
√	√	√	84.23	83.52	57.2

表2 消融实验结果(数据集2)

Table 2 Ablation experimental results (Dataset 2)

	YOLOv8	PDO-GCN	Anti Confusion	召回率/%	精确率/%	FPS
A	√			77.98	78.35	60.6
B	√	√		12.62	15.56	55.3
C	√	√	√	81.95	82.33	55.1

图12 不同数据集上PDO-GCN的检测结果

Fig.12 Detection results of PDO-GCN on different datasets

表3 不同算法的检测结果(数据集1)

Table 3 Detection results of various algorithms (Dataset 1)

算法	召回率/%	精确率/%	FPS
CrowdDet	45.94	75.63	21.2
DeepVoting	47.01	65.67	19.7
YOLOv8	74.87	81.53	61.3
YOLOv8+PDO-GCN	84.23	83.52	57.2

表4 不同算法的检测结果(数据集2)

Table 4 Detection results of various algorithms (Dataset 2)

算法	召回率/%	精确率/%	FPS
CrowdDet	78.63	71.37	19.7
DeepVoting	79.58	78.12	17.1
YOLOv8	77.98	78.35	60.6
YOLOv8+PDO-GCN	81.95	82.33	55.1

图13 数据集1上的部分实验结果

Fig.13 Partial experimental results on Dataset 1

图14 数据集2上的部分实验结果

Fig.14 Partial experimental results on Dataset 2

参考文献 20

[1]	GIRSHICK R. Fast R-CNN[C]// Proceedings of the IEEE International Conference on Computer Vision. Santiago, Chile: IEEE, 2015: 1440-1448.
[2]	REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. Advances in Neural Information Processing Systems, 2015, 1(28): 91-99.
[3]	CAI Z, VASCONCELOS N. Cascade R-CNN: delving into high quality object detection[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Salt Lake City, Utah, US: IEEE, 2018: 6154-6162.
[4]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector[C]// Proceedings of the Computer Vision-ECCV 2016: 14th European Conference. Amsterdam, The Netherlands: Springer International Publishing, 2016: 21-37.
[5]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]// Proceedings of the IEEE International Conference on Computer Vision. 2017: 2980-2988.
[6]	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. Yolov4: optimal speed and accuracy of object detection:arXiv:2004.10934[R/OL]. Ithaca, NY, US: Cornell University, 2020,(2020-04-23)[2024-05-13]. https://arxiv.org/abs/2004.10934.
[7]	TIAN Y L, LUO P, WANG X G, et al. Deep learning strong parts for pedestrian detection[C]// Proceedings of the 2015 IEEE International Conference on Computer Vision. Santiago, Washington, US: IEEE Computer Society, 2015: 1904-1912.
[8]	ZHANG S F, WEN L Y, BIAN X, et al. Occlusion-aware RCNN: detecting pedestrians in a crowd[C]// Proceedings of the 15th European Conference on Computer Vision. Munich, Cham: Springer, 2018: 657-674.
[9]	ZHANG K, XIONG F, SUN P, et al. Double anchor R-CNN for human detection in a crowd: arXiv:1909.09998[R/OL]. Ithaca, NY, US: Cornell University, 2019 (2019-09-20) [2023-10-06]. https://arxiv.org/abs/1909.09998.
[10]	CHU X G, ZHENG A L, ZHANG X Y, et al. Detection in crowded scenes: one proposal, multiple predictions[C]// Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle, WA, US: IEEE, 2020: 12211-12220.
[11]	ZHANG Z S, XIE C H, WANG J Y, et al. DeepVoting: a robust and explainable deep network for semantic part detection under partial occlusion[C]// Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, WA, US: IEEE Computer Society, 2018: 1372-1380.
[12]	枭越. M1A2主战坦克的未来发展[J]. 坦克装甲车辆, 2016(11):26-31.
	XIAO Y. Future development of M1A2 main battle tank[J]. Tanks and Armored Vehicles, 2016(11):26-31. (in Chinese)
[13]	KIPF T N, WELLING M. Semi-supervised classification with graph convolutional networks: arXiv:1609.02907[R/OL]. Ithaca, NY, US: Cornell University, 2016 (2016-09-09) [2023-10-06]. https://arxiv.org/abs/1609.02907.
[14]	WANG Y, SUN Y B, LIU Z W, et al. Dynamic graph CNN for learning on point clouds[J]. ACM Transactions on Graphics, 2019, 38(5): 1-12.
[15]	QI C R, SI H, MO K, et al. Pointnet: deep learning on point sets for 3D classification and segmentation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, US: IEEE, 2017: 652-660.
[16]	HAMILTON W, YING Z, LESKOVEC J. Inductive representation learning on large graphs[J]. Advances in Neural Information Processing Systems, 2017, 30: 1024-1034.
[17]	DUVENAUD D, MACLAURIN D, AGUILERA-IPARRAGUIRRE J, et al. Convolutional networks on graphs for learning molecular fingerprints[J]. Advances in Neural Information Processing Systems, 2015, 28: 2224-2232.
[18]	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV, US: IEEE, 2016: 770-778.
[19]	HU J G, SUN Z X, SUN Y H, et al. Progressive refinement: a method of coarse-to-fine image parsing using stacked network[C]// Proceedings of the 2018 IEEE International Conference on Multimedia and Expo. San Diego, CA, US: IEEE, 2018: 1-6.
[20]	TANG Y F, GU L Z, WANG L T. Deep stacking network for intrusion detection[J]. Sensors, 2021, 22(1): 25.