结合多尺度融合和图匹配的行人重识别
Pedestrian re-identification combining multi-scale fusion and graph matching
李冬 1张智1
作者信息
- 1. 武汉科技大学计算机科学与技术学院,湖北武汉 430065;国家新闻出版署富媒体数字出版内容组织与知识服务重点实验室,北京 100038;武汉科技大学湖北省智能信息处理与实时计算重点实验室,湖北武汉 430065;武汉科技大学 大数据科学与工程研究院,湖北武汉 430065
- 折叠
摘要
由于行人遮挡、视角变化等因素影响,传统的行人重识别并不能准确表达遮挡行人的信息.针对该问题,提出一种基于多尺度融合和图匹配的网络模型.分为提取不同尺度的特征和基于拓扑结构匹配图像两个部分,将主干网络分为两个子分支分别提取全局特征并融合多个网络层面的局部特征;使用多头注意力机制学习相邻关键点的关系,基于拓扑结构匹配图像并预测相似度结果.使用ResNet-50作为主干网络,在Occluded-Duke数据集上的Rank-1和mAP分别是64.8％和59.9％,验证该模型在遮挡行人重识别中有一定程度的准确率提升.
Abstract
Due to factors such as pedestrian occlusion and viewing angle changes,traditional pedestrian re-identification cannot ac-curately express the information of occluded pedestrians.Aiming at this problem,a network model based on multi-scale fusion and graph matching was proposed.It was divided into two parts including extracting features of different scales and matching images based on topological structure.The backbone network was divided into two sub-branches to extract global features and local features were fused at multiple network levels.The multi-head attention mechanism was used to learn the relationship between adjacent key points,and the images were matched based on the topology and the similarity result was predicted.Using ResNet-50 as the backbone network,the Rank-1 and mAP on the Occluded-Duke dataset are 64.8％and 59.9％,respectively,which verifies that the model has a certain degree of accuracy improvement in occluded pedestrian re-identification.
关键词
行人重识别/目标检测/局部特征/多尺度特征融合/图注意力机制/图匹配/卷积神经网络Key words
person re-identification/target detection/local features/multi-scale feature fusion/graph attention mechanism/graph matching/convolutional neural network引用本文复制引用
基金项目
科技创新"2030"新一代人工智能技术基金项目(2020AAA0108500)
国家自然科学基金项目(U1836118)
武汉市重点研发计划基金项目(2022012202015070)
富媒体数字出版内容组织与知识服务重点实验室开放基金项目(ZD2022-10/05)
出版年
2024
NETL
NSTL
万方数据