基于改进YOLOv3模型的车辆前方路面坑洼检测  

Detection of potholes in road ahead of vehicles based on improved YOLOv3 model

作　　者：胡均平[1] 黄强 张洪伟 向思平 宋菲菲 HU Junping;HUANG Qiang;ZHANG Hongwei;XIANG Siping;SONG Feifei(School of Mechanical and Electrical Engineering,Central South University,Changsha 410083,China)

机构地区：[1]中南大学机电工程学院,湖南长沙410083

出　　处：《传感器与微系统》2022年第12期130-133,138,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(51175518);湖南省科技型中小企业技术创新基金资助项目(12C26214305029)。

摘　　要：首先,对道路坑洼图片进行预处理以获得坑洼数据集Pothole-set;其次,将YOLOv3的激活函数修改为Mish激活函数,以提高模型准确性和泛化能力;接着,将YOLOv3的3个输出尺度进行融合以减小复杂度;然后,使用K-Means方法对坑洼数据集边界框尺寸进行聚类,同时,对坑洼数据集进行网格划分,获得最终的输出特征图;最后,将余弦退火、Mixup、标签平滑技术应用于训练过程中以提高检测精度,获得最终的坑洼检测模型YOLOv3-Pt。实验结果表明:相比于YOLOv3,YOLOv3-Pt在复杂环境下对坑洼的检测精度提升了13.99%,能够满足坑洼检测精度的需要。Firstly, the road pothole pictures are preprocessed to obtain the pothole dataset pothole-set.Secondly, the activation function of YOLOv3 is modified to the Mish activation function to improve model accuracy and generalization ability.Then, the three output scales of YOLOv3 are performed fusion to reduce complexity.K-Means method is used to cluster the bounding box size of the pothole dataset, and mesh the pothole dataset to obtain the final output feature map.Finally, Cosine annealing, Mixup, Label smoothing technology is applied in the training process to improve the detection precision and obtain the final pothole detection model YOLOv3-Pt.Test results show that compared with YOLOv3,YOLOv3-Pt improves the detection precision of potholes by 13.99 % in complex environments, which can meet the needs of pothole detection precision.

关 键 词：坑洼检测 Mish激活函数 K-MEANS聚类 余弦退火 Mixup方法 YOLOv3-Pt模型 

分 类 号：TP391.04[自动化与计算机技术—计算机应用技术]