基于多尺度特征提取与特征融合的交通标志检测

基于卷积神经网络的交通标志检测算法在对现实中复杂的交通场景图像进行交通标志检测时，难以同时解决定位和分类两项任务，并且目标检测领域相关算法所使用的公开数据集提供的图像和交通标志的种类不能满足现实交通场景中复杂的情况。建立一个新的道路交通标志数据集，在YOLOv4算法的基础上针对现实交通场景图像的复杂性和图像中交通标志尺寸差异较大的特点，设计多尺寸特征提取模块和增强特征融合模块，提高算法同时定位和分类交通标志的能力。在此基础上，对算法中不同的模块设置不同的参数进行对照实验，得到一组表现最优的参数，用于检测现实交通场景图片中的交通标志。在道路交通标志数据集上的实验结果表明，该算法相比基于卷积神经网络的同类型任务目标检测算法具有更高的检测精度，平均精度均值达到83.63%。

Traffic Sign Detection Based on Multi-Scale Feature Extraction and Feature Fusion

The traffic sign detection algorithm based on Convolutional Neural Networks (CNN) cannot easily solve localization and classification tasks simultaneously when detecting traffic signs in actual complex traffic scene images. Additionally, the images and types of traffic signs provided by the public dataset used in the related algorithms do not satisfy the situations in actual traffic.Hence, this paper presents a new dataset of road traffic signs. Subsequently, to address the complexity of actual traffic scene images and the significant variation in traffic sign sizes in the images, a multisize feature extraction module and an enhanced feature fusion module are designed based on the YOLOv4 algorithm to improve the algorithm's ability in locating and classifying traffic signs simultaneously.Different parameters are set for the different modules in the algorithm to perform comparative experiments, and a set of parameters with the best performance is obtained, which is then used to detect traffic signs in actual traffic scenes.Experimental results obtained based on the newly created dataset show that the improved algorithm achieves a mean average precision of 83.63%, which is higher than those achieved by several well-established object detection algorithms based on CNN for the same type of task.