講演抄録/キーワード |
講演名 |
2023-02-21 10:30
A Note on Traffic Sign Recognition Based on Vision Transformer Adapter Using Visual Feature Matching ○Yaozong Gan・Guang Li・Ren Togo・Keisuke Maeda・Takahiro Ogawa・Miki Haseyama(Hokkaido Univ.) |
抄録 |
(和) |
(まだ登録されていません) |
(英) |
Traffic sign recognition is a real-world task that involves many constraints and complications. Traffic sign recognition is often applied in complex systems, such as autonomous driving, with sufficient social impact. In previous studies, convolutional neural networks were used to obtain features of traffic signs from images or videos. However, the recognition accuracy directly using neural networks is not high because it is difficult to get appropriate semantic information. This paper proposes a new method for traffic sign recognition based on a vision transformer adapter model using visual feature matching. First, we obtain segmentation binary masks of traffic signs using the vision transformer adapter model. Second, the coordinates of the generated segmentation masks are computed and used to extract traffic sign images from original city images. Finally, we calculate the similarity between the removed and the template traffic sign images at the feature level to perform traffic sign recognition. Experimental results show that the proposed method achieved promising results on a real-world dataset taken in Sapporo city, Japan. |
キーワード |
(和) |
/ / / / / / / |
(英) |
Traffic sign recognition / vision transformer adapter / visual feature matching / real-world dataset / / / / |
文献情報 |
信学技報 |
資料番号 |
|
発行日 |
|
ISSN |
Online edition: ISSN 2432-6380 |
PDFダウンロード |
|