摘要
为了解决交通标志识别易受光照、遮挡和小目标影响的问题,对YOLOv5-P6算法进行改进,提出了一种新的交通标志识别算法。算法采用加权双向特征金字塔网络,提高特征提取能力,增加了跨层连接并对传递的特征进行权重调整,更好地融合道路交通标志的通道特征;使用空洞空间池化金字塔模块提取多尺度上下文信息,进一步增大感受野从而改善语义分割的效果;引入改进的跨阶段局部网络,使模块更加简洁;在训练过程中加入随机裁剪技术,并采用图像缩放、图像切变以及代数运算对检测效果不理想的类别进行实例扩充,缓解模型的过拟合问题。在TT100K数据集上应用本算法,识别精度达到90.02%,与传统的YOLOv5模型相比提高了4.72%,帧处理速率达到36.07FPS。
In order to solve the problem that traffic sign recognition was easily affected by illumination, occlusion and small targets, the YOLOv5-P6 algorithm was improved and a new traffic sign recognition algorithm was proposed. The algorithm adopts a weighted bidirectional feature pyramid network to improve the ability of feature extraction, the cross layer connection was added and the weight of the transferred feature was adjusted to better fuse the channel features of road traffic signs. The multi-scale context information was extracted by using the atrous spatial pyramid pooling module to further enlarge the receptive field and improve the effect of semantic segmentation. The improved cross-phase local network was introduced to make the module more concise. In the training process, the random cutting technology was added, and the image scaling, image shearing and algebraic operation were used to expand the examples of the categories with unsatisfactory detection effects, so as to alleviate the overfitting problem of the model. When applied to the TT100K data set, the mAP of the algorithm is 90.02%,which is 4.72% higher than the traditional YOLOv5 model, and the frame processing rate reaches 36.07FPS.
作者
李文举
张干
崔柳
沙利业
LI Wen-ju;ZHANG Gan;CUI Liu;SHA Li-ye(School of Computer Science and Information Engineering,Shanghai Institute of Technology,Shanghai 201418,China;Shanghai Precision Dosing&Weighing System Co.,Ltd,Shanghai 201108,China)
出处
《计算机仿真》
北大核心
2023年第1期149-155,共7页
Computer Simulation
基金
国家自然科学基金(61903256,61973307)。
关键词
交通标志识别
加权双向特征金字塔
空洞空间池化金字塔
数据增强
Traffic sign recognition
Weighted bi-directional feature pyramid
Atrous spatial pyramid pooling
Data augmentation
