Research On Pedestrian Re-identification Method Based On Active Learning And System Implementatio

Person re-identification aims to identify the same pedestrian through a non-overlapping multi-camera surveillance video system.In recent years,person re-identification technology has achieved a great breakthrough in the field of supervised learning.By designing models to train a large number of labeled pedestrian data,high recognition accuracy has been achieved.However,in the real scene,labeling a certain amount of pedestrian data requires a lot of time and labor cost,which makes it difficult for person re-identification technology to be applied in practice.Based on the above situation,this paper focuses on person re-identification,and proposes using active learning to solve the problem of limited training data labels.The main work is as follows:(1)An active learning method based on unsupervised multi-classification model is proposed.This method takes the unsupervised multi-classification model as the basic model,and designs a sample selection strategy with ranking uncertainty by mining the nearest neighbor relationship between samples.In the training rounds,the proposed algorithm selects the training samples with the highest uncertainty score,and makes these samples along with its nearest neighbor samples form sample pairs,which are labeled by experts and then the labeling results are further expanded according to the label propagation strategy.The labeling information will optimize multi-class pseudo labels,so that the pseudo label could be closer to the real label of the sample.Finally,experiments on three popular person re-identification datasets show that this method can effectively improve the recognition accuracy of the base model on the premise of low annotation cost.(2)In order to further reduce the labeling cost of active learning,an active learning method based on deep clustering is proposed.Based on the deep clustering model,this method designs a selection strategy based on difficult sample pairs.By analyzing the initial clustering structure,different numbers of hard sample pairs are found and labeled for each class cluster in an adaptive form,so as to determine the appropriate labeling cost according to the state of the model itself.In addition,the method also proposes to use the labeling results of hard sample pairs to guide the splitting and merging of class clusters,so as to expand the influence range of labeling information on the model pseudo labels,rather than only optimizing the pseudo labels of individual samples.Experiments on three data sets show that the manual annotation cost of this method is only about half of that of the active learning method based on unsupervised multi-classification,but the recognition accuracy has been further improved.Compared with the mainstream person re-identification methods in recent years,this method also shows excellent model performance.(3)For the person re-identification model based on active learning,an interactive person re-identification system is designed and implemented,which provides an easy operable platform for information transmission between user and model.This paper discusses the five interactive operations involved in the operation of the system: starting model training,querying model status,labeling pedestrian images and feeding back results,controlling model pause or continuation training,and visualization of search results.Finally,through the functional test of the system,the reliability of the system in each interactive operation is verified.