基于SE-YOLO-POLY模型的新型双源无轨电车智能识别  

Intelligent recognition of novel trolleybus powertrain systems based on SE-YOLO-POLY model

作　　者：董宜平 谢达 王兰 DONG Yiping;XIE Da;WANG Lan(China Electronic Technology Group Corporation No.58 Research Institute,Wuxi 214035,China;East Technologies,Inc.Wuxi,Wuxi 214072,China;Department of Fundamental Courses,Wuxi Institute of Technology,Wuxi 214121,China)

机构地区：[1]中国电子科技集团公司第58研究所,江苏无锡214035 [2]无锡中微亿芯有限公司,江苏无锡214072 [3]无锡职业技术学院基础课部,江苏无锡214121

出　　处：《南通大学学报（自然科学版）》2023年第3期72-80,共9页Journal of Nantong University(Natural Science Edition)　

基　　金：江苏省政策引导类计划(国际科技合作)-重点国别产业技术研发合作项目(BZ2018031)。

摘　　要：为解决新型的双源无轨电车的集电杆自动识别集电盒并快速并网的问题,通过改进YOLO-V4(you only look once version 4)网络模型,得到SE-YOLO-POLY(squeeze and excitation networks-you only look once version 4-POLY)网络架构。采用该网络架构,解决了由于集电盒的大小不一致、高度不一致、拍照角度不一致导致识别的集电盒出现异动的形变和尺寸变化、无法顺利并网的问题。通过SE-YOLO-POLY网络的数据集的生成、模型的设计、训练环境、实际运行反标定方式的搭建等步骤完成网络的部署。改进的模型无论在训练时间、模型大小、识别精度还是在处理速度等方面,都优于传统网络,实现了复杂环境下新型的双源无轨电车的智能并网。To address the issue of automatic recognition of collecting poles and rapid grid connection for a new type of dual-source tram without tracks,an improved YOLO-V4(you only look once version 4)network model is developed to obtain the SE-YOLO-POLY(squeeze and excitation networks-you only look once version 4-POLY)network architecture.By using this network architecture,the problem of deformations and size variations in the recognized collecting poles due to their inconsistent sizes,heights,and photographed angles,which leads to difficulties in smooth grid connection,is solved.The deployment of the network is achieved through steps such as generating the dataset of the SE-YOLO-POLY network,designing the model,setting up the training environment,and implementing the inverse calibration method during actual operation.The improved model outperforms traditional networks in terms of training time,model size,recognition accuracy,and processing speed,thereby enabling the smart grid connection of the new type of dual-source tram under complex environments.

关 键 词：YOLO-V4 SE-YOLO-POLY 目标检测与识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]