| 摘 要: 为了增强行人重识别模型(Re-identification,ReID)的遮挡感知和局部特征捕捉能力,提出一种基于特征融合的局部表征学习的方法。首先,设计遮挡样本扩充策略,通过模拟多样化的遮挡场景,有效提升模型的鲁棒性和遮挡感知能力。其次,引入局部层次编码器,在全局语义的指导下提取序列的空间相关性特征,从而增强局部特征的可鉴别性和语义完整性。实验结果显示,在Occluded-Duke和Market-1501数据集上,该方法表现出色,特别是在Occluded-Duke数据集上的rank-1达到69.2%,优于现有先进方法,提升幅度为1.3百分点,验证了该方法在提升行人重识别任务性能方面的有效性。 |
| 关键词: 行人重识别;遮挡感知;局部特征;特征融合 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金(61806071 |
|
| Re-identification of Occluded Pedestrian Based on Local Representation Learning Through Feature Fusion |
|
WANG Haowei1, YAN Gang1, GENG Shuze2
|
(1.School of Artif icial Intelligence, Hebei University of Technology, Tianjin 300401, China;
2.School of Inf ormation Technology and Engineering, Tianjin University of Technology and Education, Tianjin 300351, China)
yangang@hebut.edu.cn; 202132803002@stu.hebut.edu.cn; gengshuze@tute.edu.cn
|
| Abstract: To enhance the occlusion perception and local feature capturing ability of pedestrian re-identification (ReID) models, this paper proposes a method of local representation learning based on feature fusion. First, an Occlusion Sample Expansion Strategy(OSES) is designed to effectively improve the model's robustness and occlusion perception ability by simulating diverse occlusion scenarios. Second, a local hierarchical encoder is introduced to extract spatial correlation features of the sequence under the guidance of global semantics, thereby enhancing the discriminability and semantic integrity of local features. Experimental results on the Occluded-Duke and Market-1501 datasets demonstrate the effectiveness of the method. The rank-1 on the Occluded-Duke dataset reaches 69.2%, which outperforms the existing state-of-the-art methods by 1.2 percentage points and improves the performance of the re recognition task |
| Keywords: pedestrian re-identification; occlusion perception; local features; feature fusion |
