Pedestrian re-identification based on domain-invariant information separation guided by low-rank prior

Abstract

The invention relates to a pedestrian re-identification method based on domain-invariant information separation guided by a low-rank prior, and belongs to the field of computer vision. Embedding domain-invariant information into a dictionary learning framework, a discriminative dictionary learning model for unsupervised person re-identification across datasets is constructed. According to the low-rank prior of style information, the model can separate the domain information aliased in pedestrian image features from the domain invariant information reflecting pedestrian characteristics; at the same time, in view of the domain invariance of pedestrian attributes, the Attributes, as the link between domains, are used to construct the relationship between the source dataset and the target dataset, and reduce the domain offset between the two. Finally, the previously learned parameters are fine-tuned by a self-training strategy. Experiments show that this method approaches or even surpasses the performance of supervised non-deep learning and deep learning-based unsupervised domain adaptive person re-identification on many datasets.

Description

technical field [0001] The invention relates to a pedestrian re-identification method based on low-rank prior-guided domain-invariant information separation, and belongs to the field of computer vision. Background technique [0002] Pedestrian re-identification is a technique to search for the same pedestrian image from multiple pedestrian images under different cameras. Since this technology plays an important role in intelligent surveillance, it has attracted great attention in both academia and industry. In the actual monitoring environment, the pedestrian images captured by the camera often have low resolution. At the same time, due to the difference in viewing angle and the change of illumination, pedestrians often show strong appearance ambiguity under different viewing angles, which brings about pedestrian re-identification. great challenge. Although the performance of person re-identification based on deep learning has been significantly improved in recent years, m...

AI Technical Summary

Problems solved by technology

However, the design methods of unsupervised domain-invariant features are often difficult to mine information-rich discriminative information from the data distribution; domain adaptation methods based on adversarial learning often achieve domain alignment through adversarial learning between different domain features, but in In the process of adversarial learning, the extracted features are often shared information from samples in different domains, and the unique information between samples in different domains is lost, which easily aggravates the ambiguity between sample features

Although the method based on image style transfer is effective, it is easy to cause the label drift of the transferred image

In addition to pedestrian re-identification, the method of unsupervised domain adaptation has also received extensive attention and achieved significant research progress, but these methods often assume that the source domain and the target domain have the same or part of the same class. This assumption and The situation of pedestrian re-identification does not match

In addition, in unsupervised domain adaptation methods, the relationship between the source domain and the target domain is often a single-domain vs. single-domain problem

However, in person re-identification, both the labeled source data set and the unlabeled target data set often contain multiple camera views (each camera view can be regarded as a domain), so the method of unsupervised domain adaptation cannot directly Promoted and applied to pedestrian re-identification

Images