基于宽浅稠密网络的无人驾驶汽车交通标志牌识别

Traffic Sign Recognition of Autonomous Vehicle Based on Wide and Shallow Dense Network

以稠密网络为基础设计了交通标志牌识别模型,重点研究数据集预处理网络,利用宽浅稠密网络提取图片特征,并构建了全局平均池化分类网络。利用翻转和数据增强方法对数据集进行扩增处理,采用动态数据扩增策略使模型适应训练数据的变化,在测试集上实现了99.68%的准确率。在标志牌清晰完整和模糊不全两种情况下验证模型识别效果,结果显示,模型未出现误检和漏检情况,在图像信息被破坏的情况下,仍能以最大置信度正确地识别标志牌,识别准确度高、抗干扰能力强,具有良好的鲁棒性及泛化能力。

A traffic sign recognition model is designed based on dense network.The data set preprocessing network is the focus of the study.A Wide-shallow DenseNet is used to extract image features and global average pooling classification network is established.Data set expansion is processed through flipping and data enhancement.The dynamic data amplification strategy is adopted to adapt the model to training data changes.The accuracy rate in test set reaches 99.68%.Model recognition effect is verified under the condition of clear and fuzzy sign respectively.The results show that there is no false or missed detection.Although the image information is destroyed,the model can correctly identify the signs with the maximum confidence.It indicates that the model has high identification accuracy,strong anti-interference ability,and good robustness and generalization ability.