Cascade saccade machine learning network with hierarchical classes for traffic sign detection. (April 2021)

• Record Type: Journal Article
• Authors: Liu, Zhanwen; Qi, Mingyuan; Shen, Chao; Fang, Yong; Zhao, Xiangmo
• Is Part Of: Sustainable cities and society. Volume 67(2021)
• Publication Date: 2021-04
• Subjects: Traffic sign detection -- Attention mechanism -- Hierarchical semantic tree -- Class hierarchy -- Autonomous driving technology; Sustainable urban development -- Periodicals; Sustainable buildings -- Periodicals; Urban ecology (Sociology) -- Periodicals; 307.76
• DOI: 10.1016/j.scs.2020.102700
• Languages: English
• ISSNs: 2210-6707

Abstract

Highlights: A novel two-step cascade saccade network for traffic sign detection is proposed. Constructing a hierarchical semantic tree based on traffic labels semantic meaning. Using class hierarchy structure to transform visual features into semantic features. The proposed model can be extended to cross-domain traffic sign detection. Abstract: Traffic signs detection is one of the significant tasks for autonomous driving. It conveys notable traffic information timely to road users and maintains traffic safety in smart grid of cities. However, the size of most traffic signs is less than 0.5% of the image of traffic scene, and the uneven distribution of training samples limits the accuracy of the model. Moreover, the appearance of traffic signs within the same category of meaning always varies in shapes, colors, from one country to another. Few works have provided robust solutions to these problems simultaneously. In this paper, motivated from the property of saccade in human vision, we developed a novel architecture cascade saccade network with class hierarchy structure for traffic sign detection and domain shift problem. Experiments on Chinese traffic sign benchmarks (TT100K) demonstrated that the proposed detector achieves comparable performance with the state-of-the-art methods with an average performance improvement of 6% in precision and 14% in recall for small size, and detection time with a GPU is 0.08 s per 2048 × 2048 sized image, which can satisfy the real time requirements of driving safety applications. Moreover, proposed model can be easily extended to solve the cross-domain detection of traffic signs.