融合跨模态对齐与压缩激活机制的行人重识别算法

打开文本图片集

中图分类号：TP391.4 文献标志码：A

文章编号：1672-7010（2026）01-0011-11

Pedestrian re-identification algorithm integrating cross-modal alignment and squeeze-and-excitation mechanism

（Xiamen Institute of Software Technology，Xiamen 361024，China）

Abstract： To address the challenges of cross-modal feature extraction and matching in pedestrian reidentification—caused by variations in color，viewpoint，ilumination，and occlusion—this paper proposes an algorithm that integrates cross-modal alignment and asqueeze-and-excitation mechanism.ResNet-50 is employed to extract both modality-specificand identity-discriminative feature maps from pedestrian images. Cross-modal alignment isapplied to the modality-specific features via dynamic programming to find the shortest path for local feature alignment.Meanwhile，a squeeze-and-excitation mechanism is introduced to identitydiscriminative features： global average pooling condenses spatial information，and a two-layer full-connected network generates channel-wise weights to enhance discriminative channels and suppress noise. The fused features，combining“modalityconsistency”and“identity discriminability”，arefed back into the ResNet-50 for cross-modal pedestrian re-identification.Experimental results demonstrate that the proposed method significantly reduces intra-class distance among cross-modal features and improves feature clustering. In challenging scenarios such as heavy occlusion and cross-modal retrieval，identification accuracy improves by up to 70% compared to the baseline methods.The algorithm also exhibits enhanced robustness to occlusion and pose variations，enabling accurate pedestrian identity identification.

Key words： cross-modal alignment；squeeze-and-excitation mechanism；pedestrian re-identification; ResNet-5O；modality-specific feature map；identity-discriminative feature map

行人重识别作为智能监控、公共安全等行业的核心技术，致力于跨摄像头视角下行人身份的精准匹配[1]。（剩余11596字）