A method, device and storage medium for pedestrian re-identification based on belief learning

Abstract

The present invention provides a pedestrian re-identification method, device and storage medium based on confidence learning. The method includes: acquiring N pedestrian images from a video image resource library, randomly selecting M pedestrian images as query images, and the rest as query images Base library; use the pedestrian re-identification model to filter a certain number of pedestrian images from it and save it in a candidate training set, and mark a new person id with the number of pedestrian images; combine the candidate training set with the original training set to obtain a merged training set, and Use confidence training to find the label error of the combined training set and verify it, and then retrain the pedestrian re-identification model to obtain the re-trained pedestrian re-identification model and then update it online for pedestrian re-identification. The invention utilizes confidence learning and the existing pedestrian re-identification system to clean the massive video data under the monitoring system, thereby providing more available training data and effectively improving the generalization performance and accuracy of the pedestrian re-identification system.

Description

technical field [0001] The present invention relates to the technical field of artificial intelligence, in particular to a method, device and storage medium for pedestrian re-identification based on belief learning. Background technique [0002] With the advancement of society and technology, face recognition has increasingly become a reliable security technology. However, for most of today's cameras, their resolution often cannot meet the requirements of face recognition systems, so it is extremely necessary to use pedestrian re-identification technology that can be applied to existing monitoring systems. Pedestrian re-identification is to use image processing technology to determine whether a pedestrian under a camera appears in other cameras, so as to describe the pedestrian's moving path and achieve the purpose of cross-lens tracking. At present, the commonly used methods of person re-identification mainly include representation learning, metric learning, methods based ...

AI Technical Summary

Problems solved by technology

[0003] Pedestrians have the same structure as human faces, but unlike human faces, the structure of pedestrians is often more complex

In actual application scenarios, there may be no front face, different accessories, posture changes and occlusions, camera shooting angles, changes in indoor and outdoor environments, light differences between day and night, seasonal clothing (in winter, you may take off your coat when entering the room) ) and other unfavorable conditions and interference factors

At the same time, unlike faces that can go to the Internet to directly crawl pictures of celebrities, pedestrian re-identification is more difficult to collect and mark training data sets due to the particularity of its tasks. Existing data sets often only reach tens of thousands of people. This also hinders the further practical application of pedestrian re-identification technology, that is, the number of existing training sets is insufficient. On the other hand, there is a large amount of unlabeled data in the existing monitoring system that is difficult to use and cannot be used as a basis for pedestrian re-identification. training set use

Images