Research On Pedestrian Re-identification Method Based On Attribute Predictio

Person Re-identification(Re-ID)is a retrieval and matching technology for pedestrians in multiple camera networks,which is of great significance for the establishment of safe cities.However,person Re-ID is more challenging than other image retrieval tasks due to the influence of illumination,resolution and pedestrian posture.Pedestrian attributes,as soft biological features of biological visual monitoring,are more robust to viewpoint changes and diversity of viewing conditions,and have important potential value in person Re-ID,face recognition and identity recognition.From the perspective of pedestrian attributes,this paper fully mines the potential value of attribute information,and uses the middle level semantic information of pedestrian attributes to assist person Re-ID for single-domain and cross-domain work.The specific work is as follows:(1)A pedestrian attribute prediction method based on Task Attention Mechanism(TAM)is proposed.Based on the visual spatial and semantic relevance of attributes,this method groups attributes and improves the traditional spatial attention by referring to the multi-task learning mode.The Task Attention Mechanism(TAM)is proposed to focus on the correlation between different attribute prediction tasks and improve the performance of attribute prediction.(2)Based on the robustness of pedestrian attributes,an Augmented Pedestrian Attribute Recognition method(A-APR)is proposed by combining the attribute prediction module with the single-domain person Re-ID method.In addition,in order to improve the accuracy of attribute prediction,an attribute triplet loss is proposed to fine-tune the attribute prediction model to further improve the performance of the same-domain pedestrian re-recognition model.(3)Fully considering pedestrian attribute as a kind of general knowledge,it has certain generalization ability,that is,the consistency of attribute meaning in different domains,and carries out cross-domain generalization person Re-ID with cross-domain generalization network.The cross-domain attribute alignment strategy is proposed,which proves the rationality of cross-domain attribute prediction and provides another feasible research idea for cross-domain person Re-ID.(4)The attribute prediction method proposed in this paper shows strong performance on three common pedestrian attribute datasets: PETA,PA100 k and RAP.The A-APR method proposed in this paper also has strong competitiveness on two person Re-ID datasets:Market1501 and Duke MTMC Re-ID.In addition,we designed a cross-domain attribute experiment to verify TAM’s cross-domain attribute prediction ability,and designed two groups of cross-domain experiments using Market1501 and Duke MTMC Re-ID as source domain and target domain.The experiments show the effectiveness of the proposed method in cross-domain generalized person Re-ID task.