结合空间特征与多尺度特征融合的行人重识别方法

Pedestrian re-identification method combining spatial features and multi-scale feature fusion

针对目前行人重识别存在行人图像空间特征未对齐以及因姿态变化、遮挡使网络模型无法充分表达行人信息的问题,提出了一种结合空间特征与多尺度特征融合的行人重识别方法。首先,对空间转换网络进行改进,通过引入自约束分支自动对齐行人空间特征,解决了行人图像区域不对齐带来的空间语义信息不一致的问题;然后,从主干网络不同层级提取多尺度特征,针对性融入坐标注意力机制,并进行相互之间的融合;最后,将不同尺度特征进行融合得到了高表征能力的特征信息。在Market1501、DukeMTMC-reID和CUHK03上进行了实验,结果证明新提出的方法相比其他现有的方法拥有更高的识别准确率。

Aiming at the problem that the spatial features of pedestrian images are not aligned and the network model cannot fully express pedestrian information due to pose changes and occlusion,a method combining spatial features and multi-scale feature fusion is proposed for pedestrian re-recognition.Firstly,the spatial transformation network is improved by introducing self-constrained branches to automatically align pedestrian spatial features to solve the problem of inconsistent spatial semantic information caused by the misalignment of pedestrian image regions;Then,multi-scale features are extracted from different layers of the backbone network,targeted to incorporate coordinate attention mechanisms and fused with each other;Finally,the different scale features are fused to obtain feature information with high representation ability.Experiments are conducted on Market1501,DukeMTMC-reID and CUHK03,and the experimental results demonstrate that the newly proposed method has higher recognition accuracy compared with other existing methods.