Open AccessScale‐aware limited deformable convolutional neural networks for traffic sign detection and classification
Author(s) -
Liu Zhanwen,
Shen Chao,
Fan Xing,
Zeng Gaowen,
Zhao Xiangmo
Publication year - 2020
Publication title -
iet intelligent transport systems
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.579
H-Index - 45
eISSN - 1751-9578
pISSN - 1751-956X
DOI - 10.1049/iet-its.2020.0217
Subject(s) - convolutional neural network , computer science , traffic sign recognition , traffic sign , artificial intelligence , pyramid (geometry) , feature (linguistics) , scale (ratio) , pattern recognition (psychology) , deep learning , feature extraction , object detection , sign (mathematics) , computer vision , data mining , mathematical analysis , linguistics , philosophy , physics , mathematics , quantum mechanics , optics
Traffic sign detection and classification is a critical component of intelligent transportation systems, which is applied to inform automatic unmanned driving systems and driving assistance systems about conditions and limits of roads. Although computer vision is widely utilised in traffic sign detection, detection and recognising traffic signs globally remains a great challenge due to the variety of sign types, scale‐variance and geometric variations. To address these problems, this study proposes a region‐based deep convolutional neural network (CNN) framework for traffic sign detection and classification. Specifically, a multi‐branch sample pyramid module is proposed, which is based on multi‐branch CNNs for multi‐scaled feature exaction. A limited deformable convolutional module is then embedded into the CNN layers to learn the distorted information representation for deformation handing. Moreover, a scale‐aware multi‐task region proposal network module is applied to detect traffic signs with various scales. The whole network is trained in an end‐to‐end manner. Finally, experiments are conducted on two public detection data sets to demonstrate the effectiveness of the proposed method.