An unsupervised image video pedestrian re-identification method and system based on a migration network

Abstract

The invention belongs to the technical field of pedestrian re-identification, and discloses an unsupervised image video pedestrian re-identification method and system based on a migration network, andthe method comprises the steps: carrying out the feature extraction of an image and a video data set in a source domain through employing an improved triple network; training the generative adversarial network by using the source domain data set and the target domain training set; generating a depth feature by using the trained generative adversarial network according to the pedestrian image Itito be identified in the target training set; calculating the Euclidean distance between the depth feature of the image and the depth feature of the video in the target domain; and selecting a video which is closest to the query image, and marking a class mark which is the same as the image. According to the invention, an unsupervised method is used to eliminate a gap between an image and a video,so that the marking cost is greatly saved, and the pedestrian re-identification efficiency is improved.Through carrying out unsupervised deep learning on different modal images and videos, the cross-mode recognition efficiency is effectively improved.

Description

technical field [0001] The invention belongs to the technical field of pedestrian re-identification, in particular to a method and system for pedestrian re-identification of unsupervised image video based on transfer network, in particular to an unsupervised image video using cross-modal feature generation and transfer network of target information retention A method for person re-identification. Background technique [0002] At present, the existing technologies commonly used in the industry are as follows: [0003] The existing image-to-video pedestrian re-identification models use labeled datasets. Zhu et al. proposed a method that combines feature mapping matrices and heterogeneous dictionary pair learning. Quality image video dictionary for learning. Zhang et al. proposed a similarity learning neural network for temporal memory, including a feature representation subnetwork and a similarity subnetwork. The former uses a convolutional neural network to extract the feat...

AI Technical Summary

Problems solved by technology

In addition, for the application scenarios that require fast retrieval of surveillance videos based on a given image, marking large-scale samples requires expensive manpower and time costs, so the application of supervised person re-identification methods on these video sets challenged

[0008] (2) Existing pedestrian re-learning methods cannot handle the heterogeneity between data. These methods only use clustering or transfer learning methods to solve pedestrian re-learning between image-to-image, video-to-video and other isomorphic data. identify problem

[0013] Image and video data in the target domain usually cannot be directly used for metric learning because they do not contain labeled information.

Images