A method and system for re-identification of people changing clothes based on autoencoder network

Abstract

The invention discloses a method and system for re-identifying pedestrians who change clothes based on an autoencoder network, wherein the method includes: (1) generating a mask for the pedestrian's clothing part in the pedestrian picture; (2) using a clothing feature encoder E A Obtain the clothes feature representation vector; (3) remove the clothes part in the pedestrian picture according to the mask, and use the clothes changing picture generator G to generate the clothes changing picture; (4) build a clothes-independent feature learning network, including the feature extractor F and the picture generation (5) pair the real pedestrian pictures and the generated clothes-changing pictures in pairs, and train the feature extractor F; (6) after the training converges, input the real and marked pedestrian pictures, and use the cross-entropy loss function to The feature extractor F is fine-tuned; (7) The application of pedestrian re-identification is carried out by using the final feature extractor F. Utilizing the present invention, robust discriminative features can be learned in the scene where the appearance of pedestrians varies widely.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a method and system for re-identifying pedestrians changing clothes based on an autoencoder network. Background technique [0002] The problem of pedestrian re-identification aims to use the target person's photo as input to retrieve the historical records of the target person taken at other time points and under different cameras. Due to its wide range of applications, such as unmanned supermarkets, target person tracking, crime prevention, search for lost elderly and children, target person activity analysis, etc., the pedestrian re-identification system has rich application scenarios in real life. Therefore, the problem of person re-identification has attracted extensive attention in the field of computer vision in recent years. [0003] The problem of pedestrian re-identification is challenging, mainly due to the drastic lighting changes, pedestrian posture changes, camera ang...

AI Technical Summary

Problems solved by technology

These methods perform well in short-term person re-identification tasks, but do not perform well when applied to long-term person re-id scenarios with drastic clothing changes.

[0008] From the above two points, it can be concluded that there are currently two other key issues that hinder the commercialization of pedestrian re-identification methods: 1. Lack of labeled training data with a large number of pedestrian representation changes; 2. Lack of a method that is robust to pedestrian representation changes. The feature learning method of

Images