Guilin University of Electronic Technology
随着交通行业的发展，交通标志检测识别成为了辅助驾驶系统中最热门的研究方向之一。在实际行车道路中，交通标志具有目标小且类别繁多的特点，针对现有检测与识别算法难以同时兼顾准确度和速率的问题，提出一种YOLOv5l（you only look once version 5l）与视觉转换器（vision transformer，ViT）结合的检测与识别方法。首先采用YOLOv5l对目标进行检测，得出交通标志的位置信息，再将其输入ViT进行分类识别，其中特征连接部分引入DenseNet网络模块，来实现原始特征和卷积后特征映射的密集连接，加强特征的传递性，提高识别率。结果表明：在GTSDB和GTSRB数据集上实验效果更佳，交通标志检测速率达到200 ms，准确率达到98.78%，相比全连接层识别准确率提高了约4%。
With the development of the transportation industry, traffic signs detection and recognition had become one of the most popular research directions in advanced driving assistant system. In the actual driving road, traffic signs have the characteristics of small targets and various categories. In view of the problem that the existing detection and recognition algorithms are difficult to take into account the accuracy and speed at the same time, a detection and recognition method combining YOLOv5l and vision transformer（ViT) was proposed. Firstly, YOLOv5l was used to detect objects to obtain the location information of traffic signs. Then, the detection results were input into ViT for recognition, and the DenseNet network was added to the feature connection unit in order to densely connect the original features and the convolutional feature maps, thus enhancing the transferability of features and improving the recognition accuracy. The results show that this algorithm has better performance on GTSDB and GTSRB datasets. The detection rate of traffic signs reaches 200 ms, and the accuracy rate reaches 98.78%, which is about 4% higher than that of the fully connected network.
郭朦,陈紫强,邓鑫,等. 基于YOLOv5l和ViT的交通标志检测识别方法[J]. 科学技术与工程, , ():复制