基于FPGA的双源无轨电车的改进型YOLO-V3模型  

Improved YOLO-V3 Model for Dual-Powered Trolley Bus Based on FPGA

作　　者：董宜平 谢达 钮震 彭湖湾 贾尚杰 DONG Yiping;XIE Da;NIU Zhen;PENG Huwan;JIA Shangjie(China Electronic Technology Group Corporation No.58 Research Institute,Wuxi 214072,China;East Technologies Inc.,Ltd.,Wuxi 214072,China;eKontrol Vehicle Technology SuZhou Co.,Ltd.,Suzhou 215200,China)

机构地区：[1]中国电子科技集团公司第五十八研究所,江苏无锡214072 [2]无锡中微亿芯有限公司,江苏无锡214072 [3]凯博易控车辆科技苏州股份有限公司,江苏苏州215200

出　　处：《电子与封装》2022年第8期79-85,共7页Electronics & Packaging

基　　金：江苏省国际科技合作项目(BZ2018031)。

摘　　要：为实现双源无轨电车对集电盒的智能识别和挂载,基于第三版传统黑暗网络的主干网络单次检测(YOLO-V3)网络模型,提出以轻量化移动网络为主干网络的改进型YOLO-V3网络。通过数据集的处理、模型的设计、训练环境的搭建等完成了网络的部署,然后对模型规模、识别精度和处理速度等指标进行比较。结果显示改进型YOLO-V3网络使用更小的计算资源得到更优精度。网络部署在FPGA内部中央处理器的分散处理单元中。实车测试结果表明,改进网络明显优于其他传统网络。To realize the intelligent identification and mounting of the collector box for the dual-powered trolley bus, an improved you only look once-version 3(YOLO-V3) model with MobileNet as main stem based on YOLO-V3 of traditional Darknet was introduced. Through dealing with data sets, modelling and establishing training environment, the network was deployed and compared with traditional methods about model scale,recognition accuracy and processing speed. The simulation results showed that the proposed YOLO-V3 network had a higher precision with a lower overhead. The improved YOLO-V3 network was realized in data processing units on FPGA platform. The real trolley bus running test results showed that the improved network was better than other traditional networks.

关 键 词：YOLO-V3网络 移动网络 目标检测 FPGA 深度学习 

分 类 号：TN301.6[电子电信—物理电子学]