Research On Person Re-identification Algorithms Based On Attention Model

Person re-identification is an important research task in the field of video surveillance.As an instance-level recognition problem,person re-identification relies on discriminative features.In recent years,attention model is more and more applied to various fields of computer vision.Attention model can focus on learning the most informative part of the input signal,and can effectively improve the learning ability of the network for pedestrian feature,and can make the feature extraction network pay more attention to the discriminative features in pedestrian images.Existing algorithms tend to use only a single attention,ignoring the complementary effects of different attention network.In this article,the attention model is used as the core to research image-based person re-identification and video-based person re-identification based on the multi-attention model are proposed.First,for image-based person re-identification,this article designs a multi-attention joint learning network.This network uses Res Net50 as the backbone network,embeds the Soft attention module and the high-order attention module in the backbone network,and adds two scales of local feature extraction networks to form a multi-attention joint learning network.The network has obtained a more discriminative representation of pedestrian features.The multi-attention joint learning network was evaluated by using Market-1501 and Duke MTMC-Re ID datasets,the experimental results show that the multi-attention joint learning network proposed in this article is greatly improved over the original Res Net50 network in terms of Rank-1 and m AP indexes,and is superior to most current mainstream algorithms.On the other hand,for video-based person re-identification,this article designs a fusion multi-attention heterogeneous network.This network takes OSNet as the backbone network,in which Soft attention module and non-local attention module are embedded.At the same time,a specific local feature extraction network is added to enhance the network's learning ability for spatiotemporal information and local features of video sequences.The fusion multi-attention heterogeneous network was evaluated by using Mars and Duke MTMCVideo Re ID datasets,the experimental results show that the fusion multi-attention heterogeneous network proposed in this article has been improved to a certain extent on both Rank-1 and m AP indexes compared with original OSNet network,and it has a certain competitiveness in the state-of-art algorithms.