Abstract: To address the challenges of person re-identification (ReID) in complex scenarios such as occlusion, viewpoint variation, and pose changes, this paper proposes a spatio-temporal modeling and hierarchical feature enhancement framework. The proposed method employs a three-stage progressive feature optimization strategy to achieve a synergistic improvement in global consistency and local discriminability. First, after extracting appearance features with the backbone network, a Double Pooling Temporal Attention mechanism is introduced. This mechanism combines global average pooling and temporal average pooling to capture complementary information from sequence features, and performs spatio-temporal dependency modeling through channel–space interaction, thereby highlighting motion-related features and alleviating the problem of local information loss caused by occlusion. Second, to tackle the imbalance in body-part feature distribution, a flexible feature fusion module is designed. This module adaptively aggregates multi-part features through learnable weights, suppressing occlusion-induced noise while enhancing discriminative local representations, resulting in complementary global-local embeddings. Finally, a confidence calibration network based on residual learning is embedded before the classification layer to refine the distribution of identity prediction confidence, which effectively improves cross-camera retrieval accuracy. Comprehensive evaluations on public datasets such as Market-1501 and P-Duke-MTMC demonstrate that the proposed method achieves mAP scores of 93.2% and 86.8%, with Rank-1 accuracies of 97.4% and 95.2%, respectively. The results indicate that the integration of spatio-temporal modeling with hierarchical feature enhancement significantly improves ReID performance under challenging conditions.

Key words: Keywords: Person Re-identification, Spatio-Temporal Modeling, Flexible Feature Fusion, Feature Enhancement, Occlusion

摘要： 针对遮挡、视角变化及姿态变化等复杂场景下的行人重识别匹配困难问题，提出一种时空建模与层次化特征增强的行人重识别算法。所提算法通过三阶段渐进式特征优化框架，实现全局一致性与局部判别性的协同提升。首先，在骨干网络提取外观特征后，引入双池化时序注意力机制，该机制结合全局平均池化与时间平均池化来捕获序列特征的互补信息，并通过通道与空间交互进行时空依赖建模，从而突出运动相关特征并缓解遮挡导致的局部信息缺失问题。其次，针对人体部位特征分布不均问题，构建柔性特征融合模块，通过可学习权重自适应聚合多部位特征，抑制遮挡噪声并增强判别性局部特征，从而获得全局与局部层次化表示。最后，在分类层前设计置信度校正网络，通过残差学习优化身份预测置信度分布，提升跨摄像头检索精度。在 Market-1501和P-Duke-MTMC等公开数据集上进行系统评估，mAP分别达到 93.2%和 86.8%，Rank-1分别达到97.4%和 95.2%。结果表明，时空建模与层次化特征增强的结合能够显著提升复杂场景下的行人重识别性能。

关键词: 关键词: 行人重识别, 时空建模, 柔性特征融合, 特征增强, 遮挡