多尺度特征融合的轻量级道路坑洞检测  

Lightweight Road Pothole Detection With Multi-scale Feature Fusion

作　　者：王栋 林艺玲 谢杰 李洪帅 曾武华[1,2] 周玉珏 WANG Dong;LIN Yiling;XIE Jie;LI Hongshuai;ZENG Wuhua;ZHOU Yujue(School of Architecture and Civil Engineering,Sanming University,Samming 365004,China;Key Laboratory of Intelligent Construction and Monitoring of Engineering Structures in Fujian Provincial Universities,Sam-ming 365004,China;China Railway 24th Bureau Group Southwest Construction Co.,Ltd.,Chengdu 610052,China)

机构地区：[1]三明学院建筑工程学院,福建三明365004 [2]工程结构智能建造与监控福建省高校重点实验室,福建三明365004 [3]中铁二十四局集团西南建设有限公司,四川成都610052

出　　处：《三明学院学报》2024年第6期66-73,共8页Journal of Sanming University

基　　金：福建省自然科学基金面上项目(2022J011185);福建省中青年教师教育科研项目(JAT210421);国家级大学生创新创业训练计划项目(202111311009)。

摘　　要：道路坑洞的有效识别不仅可以提高驾乘人员的乘车舒适度,更能够为行车安全提供有力保障。针对道路坑洞检测中存在识别准确率低、检测速度慢的问题,提出一种多尺度特征融合的轻量级道路坑洞检测算法。该算法以单极多框检测器为框架,采用MobileNetV2作为特征提取骨干网络,同时将骨干网络层中3个尺度依次递减的特征层进行特征融合。实验结果表明,所提出算法的检测精度达到94.18%,相比原始算法提升7.59%,算法参数量仅为2.24 M,适用于小型嵌入式边缘计算设备,可为道路坑洞高效检测提供参考。Effective road pothole detection can not only improve the ride comfort of drivers and passengers,but also provide a strong guarantee for driving safety.To tackle the issues of low detection accuracy and slow detection speed in road pothole detection,a lightweight road pothole detection algorithm with multi-scale feature fusion is proposed.The algorithm based on the single shot mutibox detector framework,MobileNetV2 is adopted as the backbone network for feature extraction.Additionally,feature fusion is performed on three feature layers of the backbone network,which are gradually reduced in scale.The experimental results show that the algorithm achieves a detection accuracy of 94.18%,representing a 7.59%improvement over the baseline,and the number of algorithm parameters is only 2.24 M.It is suitable for small embedded edge computing devices and provides a reference for efficient road pothole detection.

关 键 词：道路坑洞 目标检测 图像特征提取 多尺度特征融合 模型轻量化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]