[1]袁 笛.基于弱监督表示学习的热红外目标跟踪[J].计算机技术与发展,2024,34(04):35-41.[doi:10. 3969 / j. issn. 1673-629X. 2024. 04. 006]
　YUAN Di.Weakly Supervised Based Representation Learning for Thermal Infrared Target Tracking[J].,2024,34(04):35-41.[doi:10. 3969 / j. issn. 1673-629X. 2024. 04. 006]

点击复制

基于弱监督表示学习的热红外目标跟踪()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年04期

页码:

35-41

栏目:

媒体计算

出版日期:

2024-04-10

文章信息/Info

Title:

Weakly Supervised Based Representation Learning for Thermal Infrared Target Tracking

文章编号:

1673-629X(2024)04-0035-07

作者:

袁 笛

西安电子科技大学 广州研究院,广东 广州 510555

Author(s):

YUAN Di

Guangzhou Institute of Technology,Xidian University,Guangzhou 510555,China

关键词:

弱监督表示学习; 主动学习; 训练样本挑选; 伪标签生成; 热红外目标跟踪

Keywords:

weakly supervised representation learning; active learning; training sample selection; pseudo-label generation; thermal infraredtarget tracking

分类号:

TP393. 0

DOI:

10. 3969 / j. issn. 1673-629X. 2024. 04. 006

摘要:

由于热红外成像技术具有更强的穿透雾、霾、雨、雪的能力,在恶劣天气条件下的成像效果几乎不受影响,使得基于热红外图像的目标跟踪任务越来越被研究者重视。 针对基于卷积神经网络的热红外目标跟踪算法在模型训练过程中需要的带有标签的数据不足的问题,提出了一种基于弱监督表示学习的方法,利用少量的标签数据及大量的无标签数据进行模型训练,从而用于热红外目标跟踪任务。 首先,利用主动学习的指导在大量无标签的数据中挑选最具有代表性的训练样本;然后,给定每个样本序列的首帧目标的真实标签,利用基础跟踪器生成该序列中其他图像帧中目标的伪标签;之后,利用带有真实标签和伪标签的训练数据进行模型训练;最后,利用训练好的模型在热红外目标跟踪算法测试数据集上进行模型测试。 实验结果表明:该方法可以在减少模型训练对标签数据需求的同时保证跟踪器的准确性。

Abstract:

Since thermal infrared imaging technology has a stronger ability to penetrate fog,haze,rain and snow,the imaging effect isalmost unaffected in bad weather conditions,which makes the target tracking task based on thermal infrared images has been paid moreand more attention by researchers. Aiming at the problem of insufficient labeled data in?

the model training of the thermal infrared targettracking algorithm based on convolutional neural network,a method based on Weakly Supervised Representation Learning?

( WSRL) isproposed,which uses a small amount of labeled data and a mass of unlabeled data for model training,so as to be used in thermal infraredtarget tracking tasks. Firstly,the guidance of active learning is used to select the most representative training samples from a large amountof unlabeled data. Then,given the ground-truth label of the target in the first frame of each sample sequence,the basic tracker is used togenerate pseudo-labels for other frames in the same sequence. Then,the training data with ground-truth labels and pseudo-labels is usedfor model training. Finally,the trained model is used to test the algorithm on the thermal infrared target tracking algorithm test data set.The experimental results show that the proposed method can ensure the accuracy of the tracker while reducing the demand for label datafor model training.

常用功能