Cross-scene continuous learning pedestrian re-identification method and device based on consistency learning

Abstract

The invention discloses a pedestrian re-identification method and device based on cross-scene continuous learning of consistency learning. The method comprises the steps of obtaining original features of data of each batch; inputting the original features into a pseudo task data conversion module, and respectively obtaining an old scene feature and a new scene feature corresponding to each original feature after scene style conversion; calculating intra-domain and inter-domain cross-scene consistency loss functions on old scene features, and performing pseudo task identity discrimination learning; calculating the pairwise similarity of the old scene feature sample and the new scene feature sample, and performing pseudo task knowledge distillation; after the new scene features are input into a classifier of the new scene, a cross entropy loss function is calculated, and identity discrimination learning is carried out; and calculating the distance between the old scene feature and the new scene feature corresponding to each sample, and carrying out cross-scene consistency learning. According to the method, the continuously updated and iterated models can be deployed in a plurality of successively learned scenes, so that the purpose of reducing the cost of model training and manual maintenance is achieved.

Description

technical field [0001] The present invention relates to the technical field of pedestrian re-identification, in particular to a pedestrian re-identification method and device based on consistent learning and cross-scene continuous learning. Background technique [0002] Most of the existing pedestrian re-identification methods are based on deep convolutional neural network (CNN) models. Training CNN models usually requires pre-collecting video data from multiple different cameras in the monitoring area of ​​a specific scene under a specific camera monitoring scene. , conduct pedestrian detection to obtain pedestrian images, and manually mark the images across cameras. And use the marked training data and the stochastic gradient descent (SGD) method to supervise the pre-training of the model, and then deploy the model to the scene. Since the shooting conditions of different video surveillance systems are usually quite different, when a model trained in a certain scene is dir...

AI Technical Summary

Problems solved by technology

[0004] The LwF and GwFReID algorithms can solve the problem of high data storage overhead to a certain extent, but because they only save a small number of samples under the old data set, only a small amount of data comes from the old scene when mixing the data of the new scene. It is not balanced, and it is easy to cause two phenomena: (1) the model is easy to overfit on the training data of the old scene, resulting in poor generalization performance in the old scene; (2) the model is more likely to be biased towards the new scene, in the Performance drops in old scenes, resulting in unbalanced recognition accuracy between old and new scenes

From the experimental results in the literature, the recognition accuracy of the LwF and GwFReID models still has a large gap with the multi-task joint training under ideal conditions.

Images