Unsupervised pedestrian re-identification method based on heterogeneous graph

Abstract

The invention discloses an unsupervised pedestrian re-identification method based on a heterogeneous graph. The method comprises the following steps: acquiring a pedestrian image and dividing the pedestrian image into a training set pedestrian image and a test set pedestrian image; constructing a camera-related heterogeneous graph for the pedestrian images in the training set; executing heterogeneous similarity propagation based on the heterogeneous graph, and efficiency mining then spatial similarity between vertexes from the heterogeneous graph; optimizing the convolutional neural network model by using heterogeneous similarity learning; and utilizing the optimized convolutional neural network model to carry out re-identification on the pedestrian image in the test set. According to themethod, the heterogeneity of the pedestrian images is fully considered on the basis of unsupervised learning, so that the defect that a traditional unsupervised learning method cannot fully consider acomplex topological structure between the unlabeled pedestrian images is effectively overcome.

Description

technical field [0001] The invention belongs to the field of intelligent security, and relates to surveillance video pedestrian re-identification analysis, in particular to an unsupervised pedestrian re-identification method based on heterogeneous graphs. Background technique [0002] Pedestrian re-identification, the main research is to accurately retrieve image frames containing a certain pedestrian from multiple cameras. Due to the existence of inconsistent lighting, inconsistent angles, occlusion and blurred faces, the research topic of person re-identification has attracted a large number of researchers to study. In these studies, although supervised training algorithms can show good results on public data sets, when these algorithms are applied to a brand new unlabeled camera network, the results are often poor and cannot be directly deployed. Therefore, how to efficiently use massive unlabeled pedestrian images for model optimization training is a challenging problem...

AI Technical Summary

Problems solved by technology

In these studies, although supervised training algorithms can show good results on public data sets, when these algorithms are applied to a brand new unlabeled camera network, the results are often poor and cannot be directly deployed. Therefore, how to efficiently use massive unlabeled pedestrian images for model optimization training is a challenging problem.

[0003] Today's mainstream unsupervised person re-identification methods are mainly based on pseudo-label and similarity methods, but these methods do not consider the complex topological structure of unlabeled data in space, and these complex topological structures can truly reflect the distribution of samples. , which can provide sufficient information to assist model optimization training

Images