Abstract: With the development of technologies such as autonomous driving and advanced driver assistance systems, the issue of traffic sign recognition (TSR) has garnered increasing attention from researchers. Currently, TSR in normal traffic environments has been well addressed, but the performance degrades significantly when there are noise interferences such as blurred or partially obstructed traffic signs. A novel TSR problem solving method combining siamese network and comprehensive feature segmentation group sparse coding is proposed to address this issue. Firstly, multiple different scale feature encodings of traffic signs are extracted, and a method using comprehensive feature encoding to represent traffic signs is proposed. Secondly, the proposed segment group sparse coding method optimizes the comprehensive feature encoding of traffic signs to enhance the model’s learning ability and improve the robustness of the encoding. Finally, a twin neural network model is constructed to train the segment group sparse coding. Due to its simple structure and fewer layers, this model is less prone to overfitting. Additionally, the proposed model requires fewer parameters, significantly improving computational speed. Experiments show that the proposed method achieves performance on par with the current SOTA models in terms of accuracy, precision, recall, and F1-score on the TT100K dataset in both the original and motion-blurred environments. Moreover, the computational parameters are reduced by 70.8%, and FPS is increased by 51.4%. In environments with partially obstructed signs, the proposed method performs notably better than the current SOTA models, especially when the occlusion rate reaches 60%, where accuracy and FPS are improved by 0.118 and 27 FPS respectively.

Key words: computer vision, traffic sign recognition, segmented group sparse coding, siamese neural network

摘要： 随着无人驾驶、辅助驾驶等技术的发展，交通标志识别（TSR）问题被更多的研究者所关注。目前，在普通交通环境下的TSR问题得到了较好的解决，但当环境中存在交通标志模糊、部分遮挡等噪声干扰时，其TSR的处理效果并不理想。针对该问题进行研究，提出了一种新颖的结合孪生网络的综合特征分段组稀疏编码的TSR问题解决方法。提取交通标志的多个不同尺度特征编码，并提出利用综合特征编码的方法来表征交通标志；通过提出的分段组稀疏编码方法对交通标志的综合特征编码进行优化，以改善模型对编码的学习能力，提高编码的鲁棒性；构建了用于分段组稀疏编码训练的孪生神经网络模型，该模型因其简单的结构和较少的层数使其不易出现过拟合问题，同时所提模型也具有较少的参数量，较大幅度提升了模型的运算速度。实验表明，所提方法在TT100K数据集原始环境、运动模糊环境中，与目前SOTA模型最好成绩相比其准确率、精确率、召回率与F1分数等评价指标相近，模型参数量减少70.8%，FPS提升51.4%；在部分遮挡噪声环境中，各指标均显著优于目前SOTA模型最好成绩，尤其在遮挡率为60%时，所提方法的准确率和FPS分别较目前SOTA模型最好成绩提升了0.118和27 FPS。

关键词: 计算机视觉, 交通标志识别, 分段组稀疏编码, 孪生神经网络