Low-resolution pedestrian re-learning method based on asymmetric mapping semi-supervised dictionary pair

Abstract

The invention belongs to the technical field of pedestrian re-identification, and discloses a low-resolution video pedestrian re-learning method based on asymmetric mapping semi-supervised dictionarypair learning. For the scene, a pair of asymmetric mappings, a pair of dictionaries of high-low-resolution videos, and a projection matrix are learned at the same time. A low-resolution video is converted into identified high-resolution characteristics by using the learnt mapping and dictionary, so that variables in each video are reduced, the gap between the high-resolution video and the low-resolution video is filled, and the discrimination between different pedestrians is clearer. According to the method, the problem that different video resolutions may exist in the recognition process, andparticularly according to a low-resolution video of a pedestrian, the pedestrian needs to be re-recognized in a high-resolution video set is effectively solved. The method can be effectively appliedto video objects with different resolutions, and the accuracy of pedestrian re-identification is remarkably improved.

Description

technical field [0001] The invention belongs to the technical field of pedestrian re-identification, in particular to a low-resolution video re-learning method based on an asymmetric mapping semi-supervised dictionary. Background technique [0002] Currently, the closest prior art: [0003] In the field of person re-identification, a class of feature representation-based methods focuses on designing a robust and discriminative feature representation for matching. For example, a discriminative feature representation model is established by utilizing category information to overcome the problem of large appearance differences between different images of the same person. Another class of methods based on matching model learning focuses on how to learn a discriminative matching model, and most of these methods adopt metric learning techniques to learn the matching model. Hirzer et al. proposed a discriminative Mahalanobis metric learning method that learns a distance metric fr...

AI Technical Summary

Problems solved by technology

[0010] (1) The application scenarios of existing pedestrian re-identification methods are mostly still images, which cannot recover or process the visual information in low-resolution pedestrian images. Some video-based pedestrian re-identification methods will extract pedestrians from video Spatio-temporal features, and use spatio-temporal features for matching in normal resolution mode, and the calculation of spatio-temporal features is also based on visual information, which also cannot handle super-resolution video re-identification well

[0011] (2) The existing semi-coupled dictionary learning method tends to deal with image super-resolution restoration tasks, but it is directly applied to super-resolution person re-identification tasks. Since the discriminator and regularization items are not designed, the learned dictionary pairs and mapping matrices may not be discriminative and efficient

[0012] (3) Since pedestrian videos often contain noise, the dictionary pair learned directly by semi-coupled dictionary learning technology cannot describe the essential relationship between high and low resolution images well

[0014] (1) Low-resolution images or low-resolution videos will cause loss of visual information and affect the performance of spatiotemporal features

It is necessary to restore the low-resolution video, and reduce the mapping error between the coding coefficients of the high-resolution and low-resolution image features as much as possible, otherwise it is difficult to effectively identify the super-resolution video

[0015] (2) The existing super-resolution restoration methods are designed to improve human visual perception, rather than machine perception that is beneficial to recognition, and cannot be directly applied to the problem of pedestrian re-identification, and the recognition effect will be suboptimal

Images