«上一篇
 文章快速检索 高级检索

 应用科技  2020, Vol. 47 Issue (2): 29-34, 43  DOI: 10.11991/yykj.201906013 0

### 引用本文

QIAN Huaming, WANG Shuaishuai, WANG Chenyu. Research on the person re-identification algorithm based on feature fusion[J]. Applied Science and Technology, 2020, 47(2): 29-34, 43. DOI: 10.11991/yykj.201906013.

### 文章历史

Research on the person re-identification algorithm based on feature fusion
QIAN Huaming, WANG Shuaishuai, WANG Chenyu
College of Automation, Harbin Engineering University, Harbin 150001
Abstract: Starting from the innovation and engineering application, a new pedestrian re-identification algorithm was proposed, which mainly solved the problem of low recognition rate and slow recognition speed in the actual pedestrian recognition system. Preprocessing the pedestrian image, using the hue, saturation, and value (HSV) spatial nonlinear quantization method to construct the color namespace, pre-identifying the human sub-regions to improve the recognition speed; extracting the HSV and direction gradient histogram of the entire oriented gradient(HOG) features as the overall feature; and sliding on the entire image of the candidate target, the color naming (CN) features and the scale-invariant local pattern (SILTP) features of the two scales are extracted in the window, getting new features by new fusion algorithms. Pedestrian re-identification is carried out on three data sets. The fusion features improve the average Rank1s of two metric learning algorithms by 2.4% and 3.3% on average. Experimental results show that the algorithm can improve the accuracy of re-identification.
Keywords: pedestrian re-identification (Re-ID)    feature extraction    nonlineaer quantization    color namespace    histogram    feature fusion    metric learning    CMC curve

1 图像预处理

2 行人重识别算法改进 2.1 特征选择

CN颜色特征属于低纬度的中级特征，对遮挡、光照等有鲁棒性。它生成一个颜色命名空间，使我们可以使用简单的语义属性来判断颜色。

2.2 HSV空间非线性量化

1）使用亮色分离策略分割HSV颜色空间，按照明度V分量值的大小分成多个色盘Vii=1，2，…，m）；

2）继续分割步骤1）得到的若干个色盘，按照饱和度S分量值的大小将每一个Vi色盘分成若干类色环Sij（1，2，…，n）；

3）在Sij空间中根据色调H分量的值就能够生成一系列的颜色块，通过人为观察，得到该颜色块属于各种颜色的概率。

2.3 行人预识别

1）计算模板图像A、B部分的主颜色，分别记为QAQB

2）计算待检测行人上身的主颜色，记为TA，当TA=QA时，说明2张图像的上身区域主颜色一致，继续执行步骤3）来计算下身区域；否则执行步骤4）。

3）计算下身区域B的主颜色，记为TB。与模板图像的主颜色进行比较，如果TB=QB，说明2张图像B区域颜色一致，我们将该图像设为待选择目标。

4）下一帧图像输入，返回步骤2）重新识别新图像。

2.4 特征提取及融合

2.5 度量学习算法

2.5.1 最大间隔近邻算法

 ${\left\| {L\left( {{x_i} - {x_{{l}}}} \right)} \right\|^2} \leqslant {\left\| {L\left( {{x_i} - {x_j}} \right)} \right\|^2} + 1$

 ${\xi _{{\rm{pull}}}}(L) = \sum\nolimits_{j \to i} \leqslant {\left\| {\left( {{{{x}}_l} - {{{x}}_j}} \right)} \right\|^2}$ (1)

 $\begin{split} &\! \!\! {\xi _{{\rm{push}}}}(L) = \\ &\! \!\! \quad \sum\nolimits_{i,j \to i} {\sum\nolimits_l {(1 - {y_{il}})} } {\left[ {1 + {{\left\| {L\left( {{x_i} - {x_{l}}} \right)} \right\|}^2} - {{\left\| {L\left( {{x_i} - {x_j}} \right)} \right\|}^2}} \right]_ + } \end{split}$ (2)

 $\xi (L) = (1 - u) \xi_{{\rm{pull}}} (L) + u \xi_{{\rm{push}}} (L)$ (3)

2.5.2 交叉二次元判别分析

 $f({{\varDelta}} ) = {{{\varDelta}} ^{\rm{T}}}({{\varSigma}} _t^{-1} - {{\varSigma}} _E^{-1}){{\varDelta}}$

 $J(\omega ) = \frac{{{{{W}}^{\rm{T}}}{{{\varSigma}} _E}{{W}}}}{{{{{W}}^{\rm{T}}}{{{\varSigma}} _t}{{W}}}}$

3 实验结果分析 3.1 数据集与评价指标

3.2 融合特征有效性分析 3.2.1 融合特征有效性分析

3.2.2 融合特征适用性分析