摘要
为解决现有基于锚框的车辆重识别模型因锚框尺寸固定导致鲁棒性差、各项预设参数十分敏感等问题,提出一种无锚框的基于改进全卷积单阶段目标检测的车辆重识别模型。基于现有FCOS模型,设计一种聚合特征金字塔网络的多层次特征模块,将特征金字塔网络最后一层的输出作为最终的重识别特征,并利用VeRi-776数据集验证了模型的有效性。实验结果表明,基于改进FCOS的车辆重识别模型在平均精度均值、Rank-1和Rank-5上均有优异表现,识别准确率分别为82.6%、96.8%和98.3%。相比现有方法,改进FCOS方法在搜索精度明显优于基于二阶段检测器的方法。
In order to solve the problems of poor robustness and sensitive preset parameters due to the fixed size of anchor frame in the existing vehicle re recognition based on anchor frame,a vehicle re-identification model without anchor frame based on improved fully convolutional one-stage object detection is proposed.Based on the classical network model,a multi-level feature module for aggregating feature pyramid network is proposed,and the output of the last layer of the feature pyramid network is used as the final re-identification feature.The validity of the model is verified on the VeRi-776 dataset.The experimental results show that vehicle re-recognition model based on improved FCOS has excellent performance in average precision,Rank-1 and Rank-5,and the recognition accuracy reaches 82.6%,96.8%and 98.3%,respectively.Compared with the classical re-identification methods,the improved FCOS method is obviously better than the method based on twostage detector in search accuracy.
作者
侯雨
穆平安
HOU Yu;MU Ping-an(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science&Technology,Shanghai 200093,China)
出处
《软件导刊》
2022年第6期57-61,共5页
Software Guide
基金
2021年学位点引导布局与建设培育项目(XWDB2021105)。
关键词
车辆重识别
全卷积单阶段目标检测
特征金字塔网络
卷积神经网络
vehicle re-identification
fully convolutional one-stage object detection
feature pyramid network
convolutional neural networks
作者简介
侯雨(1994-),男,上海理工大学光电信息与计算机工程学院硕士研究生,研究方向为图像处理;通讯作者:穆平安(1964-),男,博士,上海理工大学光电信息与计算机工程学院教授,研究方向为测试信息获取与处理、在线检测技术与装置。
