摘要
车辆再识别旨在从多个摄像机拍摄的图像中识别出同一车辆.本文提出了一种对群三元组损失函数,以特征中心点替代均值,并将对群思想和三元组损失相结合,优化了困难样本的识别.车辆再识别过程中,对群损失函数的训练过程扩大了样本规模,增加了计算量,且传统对群损失函数无法准确处理困难正样本.为此,提出了一种特征聚类对群三元组损失函数.本方法采用正样本特征聚类中心并改进了三元组损失函数的设计,从而优化了对群损失函数.在不扩增输入样本数量的同时提升了算法处理困难样本的能力.实验表明,与主流车辆再识别算法相比,本方法可有效提升车辆再识别的准确率.
Vehicle re-identification is the task of identifying the same vehicle across some images captured by multiple cameras.We propose a coupled feature clusters embedded into triplet loss dealing with hard samples.During the vehicle re-identification,the coupled clusters loss suffers from larger computation consumption caused by the extension of the sample scale and the reduction of identification accuracy.Therefore,the coupled feature clusters embedded into triplet loss is proposed.It improves the ability of the algorithm on processing hard samples in terms of selecting feature centers of positive samples based on clustering and the embedded into a triple loss.Experiments show that the algorithm effectively improves the accuracy of vehicle re-identification compared to the vehicle re-identification algorithm based on coupled clusters loss.
作者
吴燕雄
蔡建羡
滕云田
WU Yan-xiong;CAI Jian-xian;TENG Yun-tian(School of Electronic Science and Control Engineering,Institute of Disaster Prevention,Sanhe,Hebei 065201,China;Institute of Geophysics,China Earthquake Administration,Beijing 100081,China)
出处
《电子学报》
EI
CAS
CSCD
北大核心
2020年第12期2444-2452,共9页
Acta Electronica Sinica
基金
防灾科技学院教学研究与教育改革项目(No.JY2016B10)
河北省高等学校科学技术研究重点项目(No.ZD2018304)
中央高校基本科研业务费(No.ZY20180111)
国家重点研发计划项目(No.2018YFC1503801)。
关键词
车辆再识别
视觉特征
特征聚类对群损失
三元组损失
vehicle re-identification
visual appearance
coupled feature clusters loss
triple loss
