Research On Multi-modality Person Re-identification Algorithm

In recent years,with the continuous improvement of urban monitoring network,person re-identification technology has attracted more and more researchers' attention because of its potential application value.person re-identification,also known as cross mirror tracking,is one of the most popular research topics in the field of computer vision in recent years.Given a person image to be retrieved,the task of person re-identification is to retrieve all images of the person captured by cameras in different areas in a period of time.It has a wide range of applications in person tracking,behavior analysis,etc.,and it has a high accuracy under controllable(specific dataset)conditions.However,under the actual monitoring conditions,due to the influence of objective factors such as background,lighting,visual angle,hardware conditions,person re-identification becomes a challenging problem.In addition to the above limitations under visible light conditions,there are also problems caused by modality differences due to camera imaging principle.Therefore,this paper focuses on the single modality of person visible and the modality difference between visible and near-infrared in person re-identification technology.The specific contents are as follows:Aiming at the problem of person pose mismatching and the difficulty of semantic region alignment under single modality of visible modality,this paper proposes a person re-identification algorithm based on multi-granularity human semantic analysis.Firstly,a person image with key point annotation is given and retrieved;secondly,the above retrieval results are clustered to obtain the corresponding prior information;finally,the neural network modality 1 is optimized based on the prior information to obtain the final person image analysis results,that is,the four basic semantic parts of human body structure(head,upper body,thigh,calf)are completed to form multi granularity human semantic structure information.Finally,the global and local semantic features of person are obtained by softmax loss and triple loss function classification,so as to improve the accuracy of person re-identification.To solve the problem of modality difference between visible and near infrared cross modality,this paper proposes a cross modality person re-identification algorithm based on joint constraints of image and feature.It is assumed that visible and near-infrared have the same data distribution in a specific middle modality.Therefore,this paper uses a CycleGAN middle modality generation decoder based on the joint constraints of features and images to ensure that the generated persons have identity consistency.The visible modality and near-infrared modality are mapped to the same feature space with the same data distribution respectively.This reduces the difference between the two modalities of person appearance features,and improves the accuracy of person re-identification in cross modality.The contribution of this paper mainly includes the following aspects:1)To solve the problem of pose mismatching and semantic region alignment in visible single-modality,this paper proposes a person re-identification algorithm based on multi granularity human semantic analysis.Through the analysis of multi pose structure of human body,the accurate description of multi granularity person semantic segmentation area is realized,and the accuracy of visible single-modality downlink human recognition is improved.To solve the problem of modality difference between visible and near infrared cross modal downlink,this paper proposes a cross modalit person re-identification algorithm based on joint constraints of image and feature.Through a new middle modality generator between visible and near-infrared modality,the near-infrared and visible have the same data distribution in this middle modality space,reduce the difference of person appearance features in the two modalities,and improve the accuracy of person re-identification.To solve the problem of lack of cross-modality person re-identification dataset,this paper proposes a cross-modality person re-identification dataset based on cross-modality evaluation,which fills in the lack of cross-modality person re-identification experimental dataset,and further promotes the research progress of person re-identification.