Research On Person Re-identification Method Based On Multi-scale And Attention Learning

As a key link in intelligent monitoring system,person re-identification(Re-ID)has very important practical significance and is a hot research topic in the field of computer vision.Specify an image of a query person,the purpose of person Re-ID is to retrieve the person from the video or image set collected by cross cameras.Different from face recognition,person Re-ID can also complete the retrieval and tracking of pedestrians even without clear front photos,thus it has a wider range of applications.Person Re-ID is widely used in criminal investigation and tracking,intelligent human search system,intelligent security system,epidemiological investigation and many other fields,which has high research value.At present,methods based on deep learning have been effectively applied in the task of person Re-ID.However,due to the large calculation consumption,complex model structure,and the realistic problems,for example,serious occlusion,the limitation of surveillance camera views and complex background environments,it is still difficult to put models of person Re-ID into practical application.Based on deep learning,this thesis proposes a Linear-memory Attention and Group-based Multi-scale Feature Learning Network for Person Re-identification and a Person Re-ID method based on Counterfactual Attention Learning and Swin IR Transformer.The main research contents of this thesis are as follows:Aiming at the problems of large memory consumption and computational redundancy caused by the calculation of attention matrix,non-sharing of self-attention parameters and low efficiency of feature search space,a Linear-memory Attention and Group-based Multi-scale Feature Learning Method for Person Re-identification is proposed.This method is mainly composed of a Linear-memory Global Attention and a Group-level Multi-scale Feature Learning module.The Linear-memory Global Attention realizes parameter sharing by introducing external memory units.At the same time,the attention matrix is calculated by the input sequence and external memory unit,which effectively reduces the computational complexity.Considering that different pedestrians often share the same features in the real scene,the Group-level Multi-scale Feature Learning module regards the samples with shared features as a kind of grouping,which effectively improves the efficiency of feature search.In view of the problems that deep networks often ignore the learning of shallow basic information,Transformer mechanism is unable to learn the relationship between samples,and the lack of constraints on the effectiveness of attention mechanism learning,a Person Re-ID method based on Counterfactual Attention Learning and Swin IR Transformer is proposed.This method is mainly composed of an attention mechanism,a Counterfactual Attention Learning mechanism and a Swin IR Transformer.The Counterfactual Attention Learning mechanism is combined with an attention module to promote the attention module to learn more effectively.Through a shifted window attention mechanism,Swin IR Transformer realizes the learning of the relationship between samples.Experiments show that,on Prai-1581,Market1501 and Duke-MTMC datasets,the methods proposed in this thesis have achieved good performance,which improves the m AP and Rank-1 of the original method to some extent.This thesis has 28 figures,13 tables,and 95 references.