| 姬晓飞,赵帅,宋京浩,崔童.基于姿势估计和特征融合的行人重识别算法[J].电子测量与仪器学报,2024,38(4):187-194 |
| 基于姿势估计和特征融合的行人重识别算法 |
| Person re-identification algorithm based on pose estimationand feature fusion |
| |
| DOI： |
| 中文关键词: 行人重识别 深度学习 多损失函数 行人遮挡 |
| 英文关键词:person re-identification deep learning multi-loss functions person occlusion |
| 基金项目:辽宁省教育厅重点攻关项目（LJKZZ20220033）资助 |
|
|
| Author | Institution |
| Ji Xiaofei | School of Automation, Shenyang Aerospace University, Shenyang 110136, China |
| Zhao Shuai | School of Automation, Shenyang Aerospace University, Shenyang 110136, China |
| Song Jinghao | School of Automation, Shenyang Aerospace University, Shenyang 110136, China |
| Cui Tong | School of Artificial Intelligence, Shenyang Aerospace University, Shenyang 110136, China |
|
| 摘要点击次数: 58 |
| 全文下载次数: 2587 |
| 中文摘要: |
| 行人重识别在交通管理、寻找走失人口等范畴用途较广。现有算法难以处理人体姿势改变、遮挡和特征不对齐的问题，提出一种姿势引导和特征融合的行人重识别算法。所提出的算法包括3个分支，包括全局分支、基于姿势估计引导的全局分支、局部对齐分支。全局分支提取行人的全局特征，可以捕捉行人的粗粒度信息以及整体的上下文关系。基于姿势估计引导的全局分支利用姿势估计网络引导模型关注行人的全局可见区域，降低遮挡物对行人识别的干扰。局部对齐分支利用姿势估计算法构成对齐的局部特征，同时区分可见的局部区域，以降低遮挡以及姿势变化的影响。通过多分支结构，将局部特征和全局特征融合，以加强特征的多元化，增强模型的鲁棒性。最终，利用交叉熵和软边界三元损失进行模型训练。Market-1501和DukeMTMC-ReID数据集上的测试结果效验了所提算法的可行性，其间，DukeMTMC-ReID数据集的Rank-1、mAP各达成了91.2%、81.8%，具有较佳的实用性。 |
| 英文摘要: |
| Person re-identification is highly used in the areas of traffic management, searching for lost people, etc. It is hard for existing algorithms to deal with the problem of human pose change, occlusion and feature misalignment, and a pose-guided and feature-fused pedestrian re-recognition algorithm is proposed. The proposed algorithm includes three branches, including global branch, global branch based on pose estimation guidance, and local alignment branch. The global branch extracts the global features of pedestrians and captures the coarse-grained information of pedestrians. The global branch based on posture estimation guidance uses the posture estimation network guidance model to focus on the global visible area of pedestrians and reduce the interference of occlusion to pedestrian recognition. Local alignment branch uitilizes a pose estimation algorithm to establish aligned local features while distinguishing visible local regions to reduce occlusion as well as the influence of postural changes. Through a multi-branch structure, integrated local characteristics with global ones to augment feature diversity is achieved and enhanced model robustness. Finally, network training is conducted using cross-entropy and triplet loss functions. The viability of the proposed algorithm is validated by the test results on Market-1501 and DukeMTMC-ReID datasets, during which the Rank-1 and mAP of the DukeMTMC-ReID dataset reached 91.2% and 81.8%, respectively, which has a better practicality. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
