Unsupervised pedestrian re-identification method based on three-data-set cross transfer learning

Abstract

The invention discloses an unsupervised pedestrian re-identification method based on three-data-set cross transfer learning, which comprises the following steps of: training three CNNs on a big data set for image classification to obtain three pre-training models; finely adjusting the three labeled source pedestrian data sets A, B and C respectively; Utilizing the three CNNNs to respectively extract features of label-free pedestrian images in a target data set, and using K-to obtain K-values; respectively clustering the extracted features by using a neighbor clustering algorithm; screening outpicture samples which are close to a clustering center domain after clustering of the three models, and labelling the picture samples; adding the three labeled sample data into another source pedestrian data set in a crossed and alternate manner, and then finely adjusting the model; inputting a pedestrian test picture into the three trained models to obtain three feature matrixes, and performingmaximum pooling operation to obtain a unique feature of the test picture; and the Euclidean distance between the unique feature and the picture feature in the database is calculated, and the identityof the database picture with the minimum distance is the identity of the test picture.

Description

technical field [0001] The present invention relates to the field of computer vision, and more specifically, relates to an unsupervised pedestrian re-identification method based on cross transfer learning of three data sets. Background technique [0002] Pedestrian re-identification technology is a commonly used technology in the field of computer vision for detecting the identity of pedestrians in cameras in non-overlapping areas. Due to its wide application in the fields of pedestrian search, identity verification and video surveillance, this technology has received a lot of attention in recent years from more and more attention from the society. There are two main technologies in the field of pedestrian re-identification: feature expression learning and metric learning. The former is about how to let the model learn its own discriminative features to describe the appearance of pedestrians; the latter focuses on finding a suitable evaluation scale. to measure the distance...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the problem that the recognition performance cannot be improved when there are multiple label data sets in the progressive unsupervised learning method, the present invention provides a method based on three data sets An unsupervised person re-identification method based on transfer learning, which can significantly improve the person re-identification performance in the case of multiple labeled data sets

Images