摘要：

交通标志检测在智能交通领域的安全保障上具有重要作用。对于部分外观相似的交通标志在尺度变化下，现有模型难以提取它们之间的细微差异，导致标志被错误分类。此外，在复杂场景下，其他相似物体容易被误检为交通标志。为此，提出一种逐层特征细化检测模型。根据交通标志特点提出分层聚类锚框和分组损失，分层聚类锚框根据目标尺度对数据集分层并通过K-means聚类算法获取各层锚框，更好地适应交通标志灵活的尺度变化，分组损失对外观相似的类别分组并设计损失函数，指导模型学习相似交通标志间的细微差距，从而降低错误分类。在Neck层加入弱语义分割模块和特征细化模块，通过弱语义分割模块学习浅层特征上下文信息，从而分割出可信区域和非可信区域，利用特征细化模块挖掘非可信区域的上下文信息，主动学习并消除造成误检的干扰，从而降低对其他相似物体的误检。弱语义分割模块和特征细化模块结合通道注意力构建逐层细化特征金字塔，实现对多尺度特征的整体优化并提高模型精确率。实验结果表明，该方法在TT100K和GTSDB交通标志数据集上的$ \mathrm{A}{\mathrm{P}}_{50} $指标分别达到97.0%和98.6%。

关键词: 交通标志检测, 复杂场景, 分层聚类锚框, 分组损失, 弱语义分割模块, 特征细化模块

Abstract:

In an intelligent transportation system, traffic sign detection plays a crucial role. However, existing models find it difficult to extract the subtle differences between certain signs with similar appearances under scale changes, resulting in misclassification. In addition, in complex scenes, other similar objects are easily mistakenly detected as traffic signs. To this end, a layer-by-layer feature refinement detection model is proposed. Based on the characteristics of traffic signs, a hierarchical clustering anchor box and group loss are developed. The hierarchical clustering anchor box layers the dataset according to the target scale and obtains each anchor box layer by using the K-means clustering algorithm, enabling better adaptation to the flexible scale changes of traffic signs. For group loss, similar appearance categories are grouped and a loss function is designed to guide the model in learning the subtle differences between similar traffic signs, thereby reducing misclassification. A Weak Semantic Segmentation(WSS)module and Feature Refinement Module(FRM)are added in the Neck layer. The WSS module learns shallow feature context information, enabling segmentation of trusted and untrusted regions. The feature refinement module mines the contextual information of untrusted regions, actively learning and eliminating interference that causes false detections, thereby reducing false detections of other similar objects. The two modules combine channel attention to construct a layer-by-layer refinement feature pyramid, achieving overall optimization of multi-scale features and improving model accuracy. The experimental results show that the method performs well on the TT100K and GTSDB traffic sign datasets, with AP₅₀ indicators reaching 97.0% and 98.6%, respectively.

Key words: traffic sign detection, complex scenes, hierarchical clustering anchor, group loss, Weak Semantic Segmentation(WSS)module, Feature Refinement Module(FRM)