Person Re-Identification Of Cross-Dataset Domain Based On Knowledge Optimization Of Clothing Characteristics

Person Re-Identification technology refers to the use of multiple cameras in the natural scene to detect and identify whether there is a specific pedestrian under different cameras in various fields of view.Because of its great potential in the field of pedestrian tracking and intelligent security,it has been a hot research topic in academia and industry.With the deepening of the research on this technology,many excellent algorithms have been proposed and achieved satisfactory results.However,there are not many application scenarios where person re-recognition technology actually landed.A key factor is it has unavoidable bias between different domains,resulting in the poor cross-domains generalization ability of the current person re-identification model,which has greatly hindered the wide application of this technology.Although the methods based on the transfer learning and generative adversarial networks can alleviate the problem to a certain extent,these two methods have the disadvantages of low efficiency and high cost.This paper discusses the cross-scenario problem,which is the first time to use the innovative pedestrian clothing patterns to extract semantic features and transfer them to this problem,and experimental comparison results prove the effectiveness of the method.The main research contents are as follows:(1)In order to make full use of the effective information in costumes information,this paper uses a new large-scale costumes dataset for classification training.This dataset contains more than 340000 high-quality costumes images.The improved Faster R-CNN detection technology is used to detect and cut fine-grained costumes.The costumes data is constructed with hierarchical knowledge based on attribute correlation,and the costumes classification model is constructed by using multi-task learning hierarchical knowledge features.(2)This paper applies the feature transfer of costumes classification model to the cross-domains task of person re-identification.In order to verify the generalization ability of the model,a reasonable similarity matching method is selected first,and then a comparative test is carried out on three standard person re-identification datasets.The experimental results show that the model's cross-domains generalization ability is better than many mainstream algorithms.This method has the potential to reduce the expensive acquisition cost of person re-identification dataset and has a better practical application prospect.