基于异质信息对齐和重排序的跨模态行人重识别方法

Cross-modal person re-identification method based on heterogeneous information alignment and reranking

可见光图像和红外图像成像原理不同,面向可见光和红外光的跨模态行人重识别面临较大的跨模态差异,行人异质信息对齐和挖掘异常困难。为此,提出基于异质信息对齐和重排序的跨模态行人重识别方法。在异质信息对齐方面,提出一种新的异质局部信息对齐算法,通过求取行人异质局部信息距离矩阵的最短路径,实现同一行人异质局部信息的动态匹配,解决行人异质信息对齐问题;在重排序方面,提出拓展k近邻重排序算法,通过动态地拓展查询图像k近邻异质信息,实现同一行人异质信息的融合,解决行人异质信息挖掘困难问题。实验结果表明,在SYSU数据集全场景查询模式上,所提方法与AGW模型结合k近邻重排序算法相比,在评价指标mAP和Rank-1上分别提升了10.12%和8.6%。

Cross-modal person re-identification between visible and infrared light images is a challenge due to the differences in imaging principles.The alignment and mining of heterogeneous pedestrian information become difficult.To address this,we proposed a cross-modal person re-identification method based on heterogeneous information alignment and reranking.The proposed method includes a new algorithm for heterogeneous local information alignment,which dynamically matches the same pedestrian heterogeneous local information by obtaining the shortest path of the distance matrix.We also proposed an extended k-nearest neighbor reranking algorithm,which realizes the same pedestrian heterogeneous information fusion and reduces the difficulty of information mining by dynamically expanding the heterogeneous information of the query image’s k mutual nearest neighbors.The experimental results show that our method improves mAP and Rank-1 evaluation indexes by 10.12%and 8.6%respectively on the SYSU dataset compared to the AGW model combined with k mutual nearest neighbor reranking algorithm.