Scholarlink

<p>The generative adversarial networks (GAN) techniques have been widely applied to image and video field. Some of the applications are specifically related to data augmentation. However, few researchers apply GAN techniques to digital data. The target selection, whose input is digital object information, is crucial for the autonomous driving system (ADS). The correct and timely result of target selection will bring safety and comfort to the passengers. However, the bottleneck in improving target selection performance is the insufficient collection of corner cases. The conditional GAN (CGAN) techniques can effectively solve this problem. This paper designs a conditional target selection GAN (CTSGAN) model to generate time-sequence data input for target selection regarding different labels. Then, the designed CGAN is applied to generate data slots under corresponding labels. Three metrics, <i>d</i><sub>avg</sub>, <i>d</i><sub>min</sub>, and <i>d</i><sub>bounce</sub>, are created to evaluate the distribution of the data slots. Third, the training data slots are enhanced through data augmentation by the designed CGAN and applied to train the target selection model. Finally, the target selection model is deployed in an embedded system and tested under a real highway environment. The test results prove that the designed CTSGAN for data augmentation is effective and has high engineering values.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The Application of Generative Adversarial Network on Data Augmentation for Target Selection

  • 期刊论文

  • 发表日期：

    2025年6月20日

查看全文
  • Yang Liu,
  • Zihan Wang,
  • Qing Xu,
  • Keqiang Li

摘要

The generative adversarial networks (GAN) techniques have been widely applied to image and video field. Some of the applications are specifically related to data augmentation. However, few researchers apply GAN techniques to digital data. The target selection, whose input is digital object information, is crucial for the autonomous driving system (ADS). The correct and timely result of target selection will bring safety and comfort to the passengers. However, the bottleneck in improving target selection performance is the insufficient collection of corner cases. The conditional GAN (CGAN) techniques can effectively solve this problem. This paper designs a conditional target selection GAN (CTSGAN) model to generate time-sequence data input for target selection regarding different labels. Then, the designed CGAN is applied to generate data slots under corresponding labels. Three metrics, davg, dmin, and dbounce, are created to evaluate the distribution of the data slots. Third, the training data slots are enhanced through data augmentation by the designed CGAN and applied to train the target selection model. Finally, the target selection model is deployed in an embedded system and tested under a real highway environment. The test results prove that the designed CTSGAN for data augmentation is effective and has high engineering values.