摘要
行人再识别中,为了获得基于突出性颜色名称的颜色描述(SCNCD)特征对于光照变化较好的鲁棒性,提出了融合SCNCD特征和对于视角变化鲁棒性高的局部最大出现概率(LOMO)表观特征;为了获得图像的结构信息,将图像划分为多个重叠块,并提取块特征;针对神经网络容易陷入局部极小值,且收敛速度慢的问题,引入动量项。经过公用VIPeR数据库和PRID450s数据测试后,实验结果表明:融合后的特征的识别能力明显高于原特征的识别能力,且改进后的神经网络收敛速度明显提高。
In person re-identification,aiming at salient color named based color descriptor（ SCNCD） features based on salient color names are robust to illumination changes,propose to fuse the SCNCD features and the LOMO features which are robust to viewpoint changes. In order to get structure information of images,the images are divided into overlapping patches and the patch features are extracted. In order to solve the problems that neural network is easy to fall into local minimum and its convergence speed is slow,momentum term is introduced. The proposed method has been tested in the most challenging public VIPeR database and PRID450 s database,and experimental results prove that recognition abilities of the fused features are obviously higher than that of the original features,and convergence speed of the improved neural network is increased obviously.
出处
《传感器与微系统》
CSCD
2017年第8期121-125,共5页
Transducer and Microsystem Technologies
基金
国家自然科学基金资助项目(61175026)
浙江省自然科学基金资助项目(LY17F030002)
"信息与通信工程"浙江省重中之重学科开放基金资助项目(XKXL1516
XKXL1521)
关键词
行人再识别
局部最大出现频次(LOMO)特征
SCNCD特征
块特征
神经网络
pedestrian re-identification
local maximal occurrence（LOMO） features
salient color name based color descriptor（SCNCD） features
patch features
neural network
