基于深度学习的陪护机器人目标检测方法  

Object detection method of escort robot based on deep learning

作　　者：汪振耀 张礼华[1] 郑俭 WANG Zhenyao;ZHANG Lihua;ZHENG Jian(School of Mechanical Engineering,Changshan Campus,Jiangsu University of Science and Technology,Zhenjiang 212100,China)

机构地区：[1]江苏科技大学长山校区机械工程学院,江苏镇江212100

出　　处：《现代电子技术》2024年第3期51-58,共8页Modern Electronics Technique

摘　　要：为解决陪护机器人在复杂的环境下进行目标检测任务时容易出现错检、漏检等问题,提出一种基于深度学习的陪护机器人目标检测方法。首先,运用改进的WGAN网络生成高质量的图片,丰富数据集多样性,为后续YOLOv5s训练奠定基础;其次,在原YOLOv5s模型上运用Gridmask数据增强技术,提升检测模型对于复杂场景和遮挡情况下的识别能力;在Backbone部分采用C2fSE模块,并将LSTM网络与YOLOv5s网络相结合,减少了网络参数量,获得了更多的上下文信息,提高了目标检测的准确性;最后将骰子系数损失加入总的损失函数。改进后的算法与原算法相比,检测精度、召回率、mAP分别提高了6.8%、4.6%、8.3%,实验结果表明,改进后的算法在小目标物体及部分遮挡物体的检测精度、漏检等方面都有所提升,适用于家用陪护机器人的目标检测工作。In view of the false detection and missed detection when escort robots carry out object detection tasks in complex environments,an escort robot object detection method based on deep learning is proposed.The improved WGAN network is used to generate high⁃quality images to enrich the diversity of data sets and lay a foundation for subsequent YOLOv5s training.The Gridmask data enhancement technology is used for the original YOLOv5s model to improve the recognition ability of the detection model for complex scenes and occlusion.In the part of Backbone,the module C2fSE is adopted,and LSTM network and YOLOv5s network are combined to reduce the number of network parameters,obtain more context information,and improve the accuracy of object detection.The dice coefficient(DC)loss is added to the total loss function.In comparison with the original algorithm,the detection accuracy,recall rate and mAP of the improved algorithm increases by 6.8%,4.6%and 8.3%,respectively.The experimental results show that the improved algorithm has been improved in the detection accuracy and missed detection of small objects and partially⁃occluded objects,which is suitable for the object detection by household escort robot.

关 键 词：目标检测 YOLOv5s 长短时记忆网络 WGAN C2fSE 数据增强 损失函数 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TH164[电子电信—信息与通信工程]