Please wait a minute...
浙江大学学报(工学版)  2021, Vol. 55 Issue (5): 948-956    DOI: 10.3785/j.issn.1008-973X.2021.05.015
计算机与控制工程     
基于大边距度量学习的车辆再识别方法
张师林(),马思明,顾子谦
北方工业大学 城市道路交通智能控制技术北京市重点实验室,北京 100144
Large margin metric learning based vehicle re-identification method
Shi-lin ZHANG(),Si-ming MA,Zi-qian GU
Beijing Key Laboratory of Traffic Intelligent Control, North China University of Technology, Beijing 100144, China
 全文: PDF(891 KB)   HTML
摘要:

在目前的车辆再识别方法中,车辆在拍摄过程中的多视角会导致特征嵌入空间中决策边界附近样本较难区分. 针对该问题,提出通过最大化边界距离提升车辆再识别准确率. 在特征表示阶段,设计了大边界损失度量函数,可以有效处理相似车辆的混淆问题;采用入侵叛逃采样策略,可以在训练样本中找出更容易混淆的难样本以更有针对性地训练网络,并加快网络的训练速度. 在车辆检索阶段,提出基于核函数的重排序方法,可以提高车辆再识别的准确率. 在3个公共数据集上的实验结果显示,车辆再识别的准确率得到提高,同时训练和推理效率得到改善. 理论分析和实验表明,大边距度量学习通过挖掘决策边界的难样本,可以有效解决车辆再识别中的多视角问题.

关键词: 车辆再识别特征嵌入度量学习损失函数采样策略核函数    
Abstract:

In the present vehicle re-identification methods, it is hard to separate the samples near the decision boundaries in the embedding feature space because of the multi-view vehicle appearance. A large margin metric learning method was proposed to tackle the above problem. In the representation stage, a large margin loss function was presented to tackle the identity confusion issue among similar vehicles effectively. Meanwhile, an invader-defector sampler was adopted, which was used to find the hard samples in the training dataset to train the network specifically and speed up the training process. A kernelized re-rank method was adopted to further enhance the re-identification performance in the retrieval stage. Experiments on three common databases show that the proposed method can achieve a higher vehicle re-identification accuracy, and at the same time the time cost of training and the inference procedure are improved. Theoretical analysis and experiments also indicate that, the large margin metric learning method can mine the hard samples near the decision boundaries, and can solve the multi-view issue in the vehicle re-identification domain.

Key words: vehicle re-identification    feature embedding    metric learning    loss function    sampling strategy    kernel function
收稿日期: 2020-06-29 出版日期: 2021-06-10
CLC:  TP 391  
基金资助: 国家自然科学基金资助项目(61403002);北方工业大学优秀青年人才资助项目
作者简介: 张师林(1980—),男,副教授,从事人工智能研究. orcid.org/0000-0003-3034-1538. E-mail: zhangshilin@126.com
服务  
把本文推荐给朋友
加入引用管理器
E-mail Alert
作者相关文章  
张师林
马思明
顾子谦

引用本文:

张师林,马思明,顾子谦. 基于大边距度量学习的车辆再识别方法[J]. 浙江大学学报(工学版), 2021, 55(5): 948-956.

Shi-lin ZHANG,Si-ming MA,Zi-qian GU. Large margin metric learning based vehicle re-identification method. Journal of ZheJiang University (Engineering Science), 2021, 55(5): 948-956.

链接本文:

http://www.zjujournals.com/eng/CN/10.3785/j.issn.1008-973X.2021.05.015        http://www.zjujournals.com/eng/CN/Y2021/V55/I5/948

图 1  大边距度量学习整体网络结构图
图 2  不同样本在二维特征空间中的分类边界
图 3  大边距度量学习示例
图 4  入侵叛逃采样策略
数据库 方法 Rank1 Rank5 mAP
Veri-776 Res 94.35 97.30 72.10
OS 94.61 97.50 72.50
OS+BN 95.40 97.90 74.50
OS+BN+LML 95.60 97.95 75.10
OS+BN+LML+IDS 96.45 98.68 80.55
OS+BN+LML+IDS+KER 96.81 98.95 80.95
VehicleID Res 85.45 94.93 77.22
OS 85.62 95.05 77.41
OS+BN 85.92 95.12 78.45
OS+BN+LML 85.95 95.15 80.58
OS+BN+LML+IDS 88.25 97.45 86.88
OS+BN+LML+IDS+KER 88.35 98.22 88.78
表 1  在不同数据集上的消融分析
图 5  三维空间车辆嵌入特征可视化
方法 mAP Rank1 Rank5
QD-DLF[11] 61.83 88.50 94.40
Part-Reg[10] 74.30 94.30 98.70
AAVER[9] 61.20 88.97 94.70
GSTE[6] 59.40 96.24 98.97
VANet[8] 66.34 89.78 95.99
LML(IDS) 80.55 96.45 98.68
LML(IDS)+KER 80.95 96.81 98.95
表 2  Veri-776数据库上本研究方法与同类方法的准确率对比
方法 Small Medium Large
mAP Rank1 mAP Rank1 mAP Rank1
QD-DLF[11] ? 74.69 ? 68.62 ? 63.54
Part-Reg[10] 76.54 72.32 74.63 70.66 68.41 64.14
AAVER[9] 61.50 78.40 ? 75.00 ? 74.20
GSTE[6] 75.40 75.90 74.30 74.80 72.40 74.00
VANet[8] ? 88.12 ? 83.17 ? 80.35
LML(IDS) 86.60 88.25 83.55 83.42 78.38 80.33
LML(IDS)+
KER
88.75 88.31 84.45 83.85 80.68 80.75
表 3  VehileID数据集上本研究方法与同类方法的识别准确率对比
方法 Small Medium Large
mAP Rank1 mAP Rank1 mAP Rank1
DRDL[21] 22.50 57.00 19.30 51.90 14.80 44.60
FDA-Net[18] 35.10 64.00 29.80 57.80 68.41 49.40
MLSL[22] 46.30 78.40 42.40 83.00 22.80 77.50
LML(IDS) 86.47 95.94 82.85 94.61 76.63 90.75
LML(IDS)+
KER
87.45 96.28 83.35 95.15 77.35 91.85
表 4  VERI-Wild数据集上本研究方法与同类方法的识别准确率对比
图 6  车辆检索结果的可视化
图 7  车辆检索错误匹配样例分析
方法 mAP p /M e v /(ms·image?1)
QD-DLF[11] 61.83 24 ? 11.19
Part-Reg[10] 74.30 48 130 ?
AAVER[9] 61.20 48 120 ?
GSTE[6] 59.40 24 ? ?
VANet[8] 66.34 24 200 ?
LML(Res,Hard) 80.35 24 150 1.67
LML(Res,IDS) 80.50 24 110 1.67
LML(OS,Hard) 80.46 2.7 150 1.12
LML(OS,IDS) 80.55 2.7 110 1.12
表 5  本研究算法与同类算法运行效率的比较分析
1 LIU X C, LIU W, MA H D, et al. Large-scale vehicle re-identification in urban surveillance videos [C]// IEEE International Conference on Multimedia and Expo. Seattle: IEEE, 2016: 1-6.
2 LIU X C. A deep learning-based approach to progressive vehicle re-identification for urban surveillance [C]// 2016 European Conference on Computer Vision. Netherlands: Springer, 2016: 869-884.
3 ZHANG Y, LIU D, ZHA Z J. Improving triplet-wise training of convolutional neural network for vehicle re-identification [C]// 2017 IEEE International Conference on Multimedia and Expo. Hongkong: IEEE, 2017: 1386-1391.
4 ZHOU K Y, YANG Y X, CAVALLARO A, et al. Omni-scale feature learning for person re-identification [C]// 2019 International Conference of Computer Vision. Seoul: IEEE, 2019: 3701-3711.
5 HE K M, ZHANG X Y. Deep residual learning for image recognition [C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition. Nevada: IEEE, 2016: 770-778.
6 BAI Y, LOU Y H, GAO F, et al Group-sensitive triplet embedding for vehicle re-identification[J]. IEEE Transactions on Multimedia, 2018, 2385- 2399
7 SIMONYAN K, ZISSERME N. Very deep convolutional networks for largescale image recognition [C]// 2015 International Conference on Learning Representations. San Diego: IEEE, 2015: 1-14.
8 CHU R H, SUN Y F. Vehicle re-identification with viewpoint-aware metric learning [C]// 2019 International Conference of Computer Vision. Seoul: IEEE, 2019: 8281-8290.
9 PIRAZ K. A dual-path model with adaptive attention for vehicle re-identification [C]// 2019 International Conference of Computer Vision. Seoul: IEEE, 2019: 6131-6140.
10 HE B. Part-regularized near-duplicate vehicle re-identification [C]// 2019 International Conference of Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 3992-4000.
11 ZHU J Q Vehicle re-identification using quadruple directional deep learning features[J]. IEEE Transactions on Intelligent Transportation System, 2019, 21 (1): 410- 420
12 GUO H, ZHAO C. Learning coarse-to-fine structured feature embedding for vehicle re-identification [C]// 2018 the 32nd AAAI Conference on Artificial Intelligence. New Orleans: AAAI, 2018: 6853-6860.
13 ZHENG Z, ZHENG L, YANG Y A discriminatively learned cnn embedding for personre-identification[J]. ACM Transactions on Multime-dia Computing, Communications, and Applications, 2017, 14 (1): 1- 20
14 LIU X, LIU W Provid: progressive and multimodal vehicle reidentification for large-scale urban surveillance[J]. IEEE Transactions on Multimedia, 2018, 20 (3): 645- 658
doi: 10.1109/TMM.2017.2751966
15 YU R, DOU Z, BAI S Hard-aware point-to-set deep metric for person re-identification[J]. Lecture Notes in Computer Science, 2018, 196- 212
16 CORTES C, VAPNIK V Support-vector networks[J]. Machine Learning, 1995, 20 (3): 273- 297
17 ZHONG Z, ZHENG L. Re-ranking person re-identification with k-reciprocal encoding [C]// 2017 International Conference of Computer Vision and Pattern Recognition. Honolulu: IEEE, 2019: 3652-3661.
18 LOU Y, BAI Y. Veri-wild: a large dataset and anew method for vehicle re-identification in the wild [C]// 2019 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2019: 3230-3238.
19 LIU H, TIAN Y. Deep relative distance learning: tell the difference between similar vehicles [C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2019: 2167–2175.
20 MAATEN L, HINTON G Visualizing data using t-SNE[J]. Journal of Machine Learning Review, 2008, 9: 2579- 2605
21 LIU Y, TIAN Y. Deep relative distance learning: tell the difference between similar vehicles [C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2019: 2167-2175.
[1] 刘芳,汪震,刘睿迪,王锴. 基于组合损失函数的BP神经网络风力发电短期预测方法[J]. 浙江大学学报(工学版), 2021, 55(3): 594-600.
[2] 秦洪远, 刘一鸣, 黄丹. 脆性多裂纹扩展问题的近场动力学建模分析[J]. 浙江大学学报(工学版), 2018, 52(3): 497-503.
[3] 贾松敏,卢迎彬,王丽佳,李秀智,徐涛. 分层特征移动机器人行人跟踪[J]. 浙江大学学报(工学版), 2016, 50(9): 1677-1683.
[4] 赵振,张树有. 基于分步回归的低压断路器分断电流求解技术[J]. J4, 2012, 46(11): 1943-1952.
[5] 厉小润, 伍小明, 赵辽英. 非监督的高光谱混合像元非线性分解方法[J]. J4, 2011, 45(4): 607-613.