Research And Application Of Multi-modality Person Re-identification

Person re-identification is an important research direction in the field of intelligent security.It aims to use query images to retrieve related images in an image database,which can be regarded as an image retrieval technology.Most of the current work of person re-identification focuses on visible-modality scenes.In real scenes,the visible camera can only take images of pedestrians in the daytime or scenes with sufficient light conditions and has poor adaptability to nighttime or scenes with poor light conditions.With the advent of infrared cameras,it is possible to obtain infrared images of pedestrians under poor light conditions.Therefore,some researchers begin to pay attention to the person re-identification problem under infrared-modality scenes and cross-modality scenes.Obviously,due to the single-channel characteristic of infrared images,it lacks a lot of information compared to visible images.Therefore,person re-identification under infrared-modality scenes and cross-modality scenes is more difficult than that in visible-modality scenes.In this paper,the person re-identification problem is studied and divided into visible-modality(RGB-RGB),infrared-modality(IR-IR),and cross-modality(RGB-IR)scenes.First of all,this paper categorizes and summarizes some popular algorithms of three modality scenes in detail,explains their basic principles,and summarizes the relevant results.Among them,visible-modality scenes are divided into full-supervised scenes,unsupervised scenes,and semi-supervised scenes.Second,based on the infraredmodality scenes,considering there is no public dataset currently,this paper proposes an infrared-modality person re-identification dataset called NJU-IR for the first time which mainly contains 30,385 images of 1,564 pedestrians,and introduces the constructionprocess of the dataset from the perspective of person detection,person tracking,etc.A large number of existing visible-modality methods are experimented on the dataset,and a new benchmark is proposed for the dataset,which also illustrates the challenge and research value of the dataset.Third,based on the cross-modality scenes,considering that the shortcomings in feature space representation and loss function design of some existing methods,this paper proposes an improved two-stream network algorithm based on grayscale features.This method inherits the network structure and the hetero-center loss based on local features in the TSLFN framework and adds grayscale conversion module and the global feature expression into the framework,which improves the feature representation ability of the shared modality space and solves the problem that the global feature is not considered.The comparison experiment proves that it achieves a significant improvement effect on the SYSU-MM01 and Reg DB datasets compared to other methods,and the ablation experiment also proves the effectiveness of the method.Finally,for the above three modality scenes and some corresponding popular algorithms,the paper implements a multi-modality person re-identification system by the web application framework Django based on Python language.It is used to realize the retrieval and re-identification tasks of the specified pedestrians and visualization in the above three modality scenes,which takes the person re-identification technology into practical applications.