Scholarlink

We propose the use of vision transformer models for speed estimation based on audio captured from the roadside. Rather than processing audio as one-dimensional signals, we trained the models using Mel-spectrograms from each recording session, incrementally augmenting the dataset to efficiently train the vision transformer models. In this paper, we compare the performance of various state-of-the-art vision transformer models, such as CrossFormer and SwinViT, against convolutional models like ResNet and EfficientNet. Our results demonstrate that vision transformers perform on par with, or even outperform, CNN models in this task, with CrossFormer achieving an R2 score of 92.5%.

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

Speed Estimation from Audio Using Vision Transformers

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Salim Abdelaziz,
  • Salim Ziani

摘要

We propose the use of vision transformer models for speed estimation based on audio captured from the roadside. Rather than processing audio as one-dimensional signals, we trained the models using Mel-spectrograms from each recording session, incrementally augmenting the dataset to efficiently train the vision transformer models. In this paper, we compare the performance of various state-of-the-art vision transformer models, such as CrossFormer and SwinViT, against convolutional models like ResNet and EfficientNet. Our results demonstrate that vision transformers perform on par with, or even outperform, CNN models in this task, with CrossFormer achieving an R2 score of 92.5%.