疑似画像を用いた事前学習と実画像を用いた本学習による夜間道路標識検出

カテゴリ: 論文誌(論文単位)

グループ名: 【C】電子・情報・システム部門

発行日: 2021/09/01

タイトル(英語): Pre-training using Pseudo Images and Fine-tuning using Real Images for Nighttime Traffic Sign Detection

著者名: 山本　雅也（静岡大学総合科学技術研究科），大橋　剛介（静岡大学総合科学技術研究科）

著者名(英語): Masaya Yamamoto (Shizuoka Univercity), Gosuke Ohashi (Shizuoka Univercity)

キーワード: 道路標識，深層学習，物体検出，事前学習，fine-tuning，ADAS　　traffic sign，deep learning，object detection，pre-training，fine-tuning，ADAS

要約(英語): In this study, a large amount of pseudo images for pre-training is created to achieve high accuracy even with the limited amount of ground truths in nighttime traffic sign detection. First, a large amount of pseudo images is created which does not require manual process for acquiring ground truths by randomly changing parameters of traffic sign template images with conditions based on the analysis of existing ground truths. Next, the model initialized by ImageNet weights is pre-trained with the created large amount of pseudo images, and finally the pre-trained model is fine-tuned with real images. By using a large amount of pseudo images for pre-training, suitable weights for pseudo nighttime traffic sign detection are obtained. Finally, the pre-trained weights can be adjusted for real nighttime traffic sign detection even with the limited amount of data by fine-tuning the model with real images. Experimental results show that as the detection accuracy of proposed fine-tuned model is improved compare with the model without pre-training by pseudo images, pseudo images can be used for effective training of a model.

本誌: 電気学会論文誌C（電子・情報・システム部門誌） Vol.141 No.9 （2021） 特集：知能メカトロニクス分野と連携する知覚情報技術

本誌掲載ページ: 969-976 p

原稿種別: 論文／日本語

電子版へのリンク: https://www.jstage.jst.go.jp/article/ieejeiss/141/9/141_969/_article/-char/ja/