Visual pedestrian re-recognition method based on sparse graph similarity migration

Abstract

The invention provides a visual pedestrian re-recognition method based on sparse graph similarity migration. The method comprises the steps: extracting features of each pedestrian image and a query image in a pedestrian image database through employing the same trained deep convolutional network, and representing the features through feature vectors; calculating the similarity of any two pedestrian images through the feature vectors, and constructing a database image dense association graph; performing sparse constraint on the database image dense association graph to obtain a database image sparse graph; using an energy minimum random walk model to take a value for calculating the similarity between the query image and the pedestrian image as an energy random migration in a database image sparse graph, taking a stabilized energy value as a consistency score of the query image and the pedestrian image, and sorting the database image sparse graph based on the consistency score to obtain a database image sparse graph; and returning the pedestrian image with the highest score. By the adoption of the scheme, the retrieval precision and speed of visual pedestrian re-recognition are improved.

Description

technical field [0001] The invention relates to the field of computer algorithms, and can be applied to reordering learning of pedestrian re-identification tasks, in particular to a visual pedestrian re-identification method based on sparse graph similarity migration. Background technique [0002] Pedestrian re-identification can help people quickly find specific people. With the help of the current increasing number of urban surveillance cameras, it can maintain social order, promote social public safety and even national security. Due to its important role and outstanding value in the security field, person re-identification has become an important research direction in the computer vision research community. Researchers have been trying to establish a powerful pedestrian re-identification model, which can expressly establish the relationship between a specific pedestrian portrait and the pedestrian picture in the real picture under the surveillance camera, and then achiev...

AI Technical Summary

Problems solved by technology

[0005] The current re-identification method can greatly improve the retrieval accuracy of pedestrian re-identification, but there are also problems of high algorithm complexity and low computational efficiency

For example, the method based on query expansion needs two feature measurement learning and one feature fusion, the k-nearest neighbor method needs to traverse all images to determine the images that are neighbors, and the manifold learning method needs to establish dense associations for a large number of labeled data. Graph for label propagation learning, although offline learning can be used, but the huge association graph limits its computational efficiency

Images