基于姿态估计与特征相似度的行人重识别算法

Person Re-Identification Based on Pose Estimation with Feature Similarity

针对行人重识别场景复杂引起的局部特征不对齐,以及在背景杂乱情况下难以提取出具有不变性行人特征的问题,提出一种基于人体姿态估计算法和相似度矩阵引导的多尺度融合网络.网络引入姿态估计算法构造对齐的行人特征,通过多分支结构将低层局部特征和高层全局特征进行融合提升网络的表征能力;此外特征相似度矩阵将全局特征分割出相似度引导的背景、前景分支,再利用区域级的三元损失函数,提取出对复杂背景鲁棒的行人特征.在Market-1501、DukeMTMC-ReID、CUHK03和MSMT17四个主流数据集上的实验结果表明,本文提出的方法均能达到甚至超过当前主流算法的水平.在最具挑战的MSMT17数据集中,与目前精度领先的算法相比,首次命中精度提高了 1.4个百分点,平均精度均值提高了 3.4个百分点.

This paper proposes a part aligned multiscale fusion network,based on a human pose estimation algorithm and similarity matrix to address the misalignment problem of local features caused by complex person reidentification scenes and difficulty in extracting invariant person features in cluttered backgrounds.The proposed network introduces poseestimation algorithms to construct aligned local features and integrates lowlevel local features and highlevel global features through a multibranch structure.In addition,the feature similarity matrix divides the global features into the similarityguided background and foreground branches and uses the regionallevel triplet loss to extract person features robust to complex backgrounds.Extensive experiments are conducted for four datasets(Market-1501,DukeMTMCReID,CUHK03,and MSMT17).The proposed method achieves stateoftheart performance.In particular,it improves the accuracy of first hit accuracy by 1.4 percentage points and mean average precision by 3.4 percentage points on the most challenging MSMT17 dataset.