期刊
NEURAL COMPUTING & APPLICATIONS
卷 34, 期 10, 页码 8241-8252出版社
SPRINGER LONDON LTD
DOI: 10.1007/s00521-022-06903-4
关键词
Attention mechanism; Pose estimation; Multitask network; Person re-identification
资金
- National Natural Science Foundation of China [61866004, 61966004, 61962007]
- Guangxi Natural Science Foundation [2018GXNSFDA281009, 2019GXNSFDA245018, 2018GXNSFDA294001]
- Research Fund of Guangxi Key Lab of Multi-source Information Mining and Security [20-A-03-01]
- Innovation Project of Guangxi Graduate Education [JXXYYJSCXXM-2021-013]
- Guangxi ''Bagui Scholar'' Teams for Innovation and Research Project
In this study, a pose-driven attention fusion mechanism is proposed to address the occlusion problem in pedestrian re-identification. By fusing pose-driven key points and spatial attention, the interference caused by occlusion is eliminated and the identification granularity is improved by matching local features. Experimental results demonstrate that the proposed method achieves competitive performance.
Pedestrians are often occluded by various obstacles in public places, which is a big challenge for person re-identification. To alleviate the occlusion problem, we propose a Pose-drive Attention Fusion Mechanism (PAFM) that jointly fuses the discriminative features with pose-driven attention and spatial attention in an end-to-end framework. To simultaneously use global and local features, a multi-task network is constructed to realize multi-granularity feature representation. After anchoring the region of interest to the un-occluded spatial semantic information in the image through the spatial attention mechanism, some key points of the pedestrian's body are extracted using pose estimation and then fused with the spatial attention map to eliminate the harm of occlusion to the re-identification. Besides, the identification granularity is increased by matching the local features. We test and verify the effectiveness of the PAFM on Occluded-DukeMTMC, Occluded-REID and Partial-REID. The experimental results show that the proposed method has achieved competitive performance to the state-of-the-art methods.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据