面向路面抛洒物检测的多任务学习算法研究  

Research on Multi-task Learning Algorithms for Pavement Spills Detection

作　　者：井晶[1,2] 赵广明 赵作鹏 JING Jing;ZHAO Guangming;ZHAO Zuopeng(Jiangsu United Vocational and Technical College,Xuzhou Finance and Economics Branch,Xuzhou 221008,China;School of Computer Science and Technology,China University of Mining and Technology,Xuzhou 221116,China)

机构地区：[1]江苏联合职业技术学院徐州财经分院,江苏徐州221008 [2]中国矿业大学计算机科学与技术学院,江苏徐州221116

出　　处：《许昌学院学报》2024年第5期111-117,共7页Journal of Xuchang University

基　　金：国家自然科学基金(61976217)。

摘　　要：针对现有检测算法在定点抛洒物检测方面的不足,提出一种基于多任务学习和图像处理的路面抛洒物检测算法.首先,为YOLOv6s设计一种融合多层语义特征的语义分割头,并为目标检测分支和语义分割分支设计损失函数,进行多任务学习;其次,结合语义分割分支分割连续性的优势以及图像处理的方法,提取完整路面区域;最后,引入结合IoU和质心位置的抛洒物区域识别与排除方法,实现非抛洒物的排除和最终抛洒物区域提取.实验结果表明,在BDD100k数据集上,改进算法在车辆目标检测和可行驶区域分割上分别达到77.8%的平均精度值(mAP)和91.5%的平均交并比(mIoU).To address the shortcomings of existing detection algorithms in detecting road-spill objects,a novel approach based on multi-task learning and image processing for road-spill object detection is proposed.Firstly,a semantic segmentation head that integrates multi-layer semantic features is designed for YOLOv6s,and loss functions are devised for both the object detection and semantic segmentation branches to enable multi-task learning.Secondly,by leveraging the advantage of continuity in the semantic segmentation branch,coupled with image processing techniques,the complete road area is extracted.Finally,a method for identifying and excluding spilled object regions is introduced,which combines IoU and centroid position,enabling the exclusion of non-spilled objects and the final extraction of spilled object regions.Experimental results demonstrate that on the BDD100k dataset,the improved algorithm achieves a mean Average Precision(mAP) of 77.8% for vehicle object detection and a mean Intersection over Union(mIoU) of 91.5% for drivable area segmentation.Additionally,spilled objects of various scales can be accurately extracted.The algorithm can be directly deployed on vehicle-mounted equipment for existing road damage inspection tasks.

关 键 词：多任务学习 抛洒物检测 YOLOv6s 图像处理 语义分割 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]