行人重识别模型的多任务损失设计  

Multi-task loss design of person re-identification model

作　　者：白宗文[1] 张哲 BAI Zongwen;ZHANG Zhe(School of Physics and Electronic Information,Yan’an University,Yan’an 716000,China)

机构地区：[1]延安大学物理与电子信息学院,陕西延安716000

出　　处：《西安科技大学学报》2024年第2期400-408,共9页Journal of Xi’an University of Science and Technology

基　　金：国家自然科学基金项目(62266045)。

摘　　要：行人重识别是一项利用计算机视觉技术判断图像中是否存在特定行人的任务。为研究Re-ID模型使用身份标签不能有效地学习不同行人之间的相似局部外观问题,提出了一种基于多任务损失的Re-ID方法。首先,通过主干网络提取全局特征以及局部特征,借助姿态估计算法检测行人身体部位,将身体部位的特征与局部特征组进行融合形成人体姿态引导特征;其次,通过多任务损失方法指导模型对人体姿态引导特征以及全局特征进行优化,从而增强模型对遮挡以及不具有区分性局部外观的鲁棒性。结果表明:多任务损失方法在Occluded-Duke、Market 1501和DukeMTMC-reID数据集上的mAP/Rank-1的精度分别达到了59.7%/67.9%,88.4%/94.9%和80.6%/89.9%。为避免训练集与测试集数据之间分布的差异性导致预训练模型产生次优检索结果的问题,提出了一种基于图卷积网络的重排序方法,该方法利用图卷积算子在图上将行人的最近邻特征传播,从而优化了每个图像的表示,以获得更优的检索结果。Person re-identification is a task that utilizes computer vision technology to discern the presence of specific pedestrians within images.In tackling the problem of Re-ID models struggling to effectively learn similar local appearance between different pedestrians when employing identity labels,a approach based on multi-task loss has been introduced.Initially,global and local features are extracted via the backbone network,with a pose estimation algorithm utilized to detect pedestrian body parts.Then,integrate the features of body parts with local features to form human pose-guided features.Subsequently,through a specially designed multi-task loss methodology,the model is guided to optimize both human pose-guided features and global features,thereby fortifying its robustness against occlusion and non-discriminative local appearances.The results indicate that this approach achieves precision rates of 59.7%/67.9%,88.4%/94.9%,and 80.6%/89.9%for mAP/Rank-1 across the Occluded-Duke,Market 1501,and DukeMTMC-reID datasets,respectively.To mitigate the impact of distribution discrepancies between training and testing datasets on the performance of pre-trained models,a re-ranking strategy based on graph convolutional networks is proposed.By leveraging graph convolution operators,this method propagates nearest neighbor features of pedestrians on the graph to refine the representation of each image,thereby enhancing retrieval outcomes.

关 键 词：行人重识别 姿态估计算法 多任务损失 图卷积算子 重排序 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]