| Person re-identification mainly studies how to identify specific pedestrians shot by different surveillance cameras from different perspectives.In recent years,with the continuous development of smart city construction,person re-identification technology has attracted more and more attention.The existing person re-identification methods are still faced with the challenges of angle change,posture change,light change and pedestrian occlusion.In order to effectively solve the above problems,this paper,based on deep convolutional neural network,introduces attribute recognition and attention mechanism to enhance the ability of neural network to obtain high-level semantic information and spatial information,so as to improve the performance of the whole person re-identification system.In this paper,an attribute recognition fusion convolutional neural network is first proposed.In order to obtain better high-level semantic information,this network introduces attribute information and innovatively differentiates attribute space and pedestrian identity space by adding feature dimension,so that attribute recognition task and person reidentification task can promote each other well.Then this paper proposes an attention fusion convolutional neural network.The innovation of this network lies in the feature attention block,which uses residual structure to fuse spatial attention and channel attention,so that the features obtained from the network can highlight the spatial position of pedestrians.In order to effectively fuse feature attention block,the network also constructs branches by using features of the backbone network in different stages and embes the feature attention block in the branches.However,attention to the whole pedestrian does not enable the person re-identification system to distinguish the subtle attributes of different pedestrians.In this paper,the existing problems are overcome by adding attribute recognition into the network that integrates the attention mechanism,and an attention attribute-identity feature fusion network is finally proposed.In this network,the attribute recognition branch is constructed to combine features of different stages,and the feature attention block is used to mine high-level semantic information and spatial information.The output of the network will fuse the characteristics of branches and backbone networks to form the final pedestrian feature representation.The experiments on the Market-1501 and DukeMTMC-reID show that the accuracy of the proposed three network has very big enhancement compared with the baseline network,and the attention attribute-identity feature fusion network outperforms most of the research methods at present,which reaches 93.9% and 84.6% respectively in Rank1,and obtains 81.7% and 70.7% in mAP,respectively.Meanwhile,the average accuracy of attribute recognition also reaches 92.4% and 91.3% on Market-1501 and DukeMTMC-reID,respectively. |
