MC-Net: multi-scale feature fusion and cross-level information interaction network for traffic sign detection

Date

2023

Authors

Yu, Z.
Cheng, D.
Zhang, W.
Chen, J.
Zhang, S.

Editors

Ceballos, C.

Advisors

Journal Title

Journal ISSN

Volume Title

Type:

Conference paper

Citation

Proceedings - International Conference on Tools with Artificial Intelligence, ICTAI, 2023 / Ceballos, C. (ed./s), pp.841-848

Statement of Responsibility

Conference Name

The 35th IEEE International Conference on Tools with Artificial Intelligence (ICTAI) (6 Nov 2023 - 8 Nov 2023 : Atlanta, USA)

Abstract

Traffic sign detection is an important topic in autonomous driving and intelligent transportation, as it is widely applied in real-life scenarios to detect crucial road information for autonomous devices. However, existing detection methods face challenges due to significant variations in the scale of target objects and limited preservation of contextual information. To address these challenges, we propose MC-Net, a novel traffic sign detection network that enhances the model’s receptive field and enriches contextual information features. To overcome the sensitivity of detectors on the target scale variation, we first introduce the Multi-Scale Feature Fusion (MFF) module. By incorporating dilation convolution and group convolution, the MFF module effectively expands the network’s receptive field and reduces sensitivity to scale variations. Additionally, we incorporate a cross-layer information interaction (CLII) module into the MC-Net model, which facilitates the extraction of feature information across diverse network layers and effectively addresses the issue of contextual information loss. Experimental results on two real-world datasets (TT100K and STSD), demonstrate that MC-Net significantly improves detection efficiency compared to existing methods, achieving mAP of 85.7% and 94.8% respectively.

School/Discipline

Dissertation Note

Provenance

Description

Access Status

Rights

Copyright 2023 The Institute of Electrical and Electronics Engineers Access Condition Notes: Accepted manuscript available open access

License

Grant ID

Call number

Persistent link to this record