With the rapid improvement of computer processing ability and development of deep learning technology,person re-identification(ReID)has become a key technology of social intelligent security,and ReID has gradually moved from algorithm research to practical applications,such as criminal investigation,surveillance,security,and other fields.However,the data acquisition in the person ReID scene is mostly passive,that is,people do not cooperate with the camera,so the quality of data is worse than that of face data.Moreover,the sensitivity of ReID application field needs a more reliable prediction of the model.Therefore,providing an accurate and reliable model has become an inevitable development direction.The uncertainty estimation method can usually improve the reliability and robustness of the deep model,which can provide great help to the practical application of ReID.Based on the above background,this thesis focuses on the application of data uncertainty estimation in person ReID.The main contributions are summarized as follows:1.This thesis proposes a person ReID model based on data uncertainty estimation method.According to the characteristics of person data,this thesis conducts uncertainty modeling and analysis on local features,which introduces uncertainty components by designing uncertainty branches in the model and embedding local person features.Moreover,the original feature embeddings component(mean)and the uncertainty(variance)are introduced into the feature and learned by the model,which also gives the person ReID model the ability to estimate both local and global uncertainties and improves the robustness of the extracted noise sample features.Comprehensive experiments are conducted on the generated noise dataset,and the results show that our method achieved the best performance.2.This thesis designs a new uncertainty estimation loss function for the ReID task.The proposed loss function constrains uncertainty and features at the same time,which not only solves the problem that the uncertainty branch is difficult to converge in the training phase but also avoids the problem of bias and trivial solutions.By comparing the model with the proposed loss function on the Market1501 and DukeMTMC-reID datasets,the results show that the performance(Rankl,mAP)is further improved.3.This thesis also proposes an uncertainty-based person data quality assessment method.As a non-referenced objective quality assessment method,this method has obvious advantages over the traditional methods.Through the artificial perception quality evaluation test and the quality filtering experiment on the iLIDS-VID and MARS datasets,we conduct experiments from both subjective quality evaluation and objective quality evaluation and compare with another non-referenced quality evaluation method QAN.All the results show the effectiveness of the uncertaintybased person data quality assessment method and also indicate that uncertainty estimation can be effectively applied in the person ReID task. |