Research On Cross-modality Re-identification Based On Deep Learning

In recent years,person re-identification has been widely used in video surveillance,network criminal investigation and other fields,mainly to solve the retrieval problem of the same pedestrian between multiple cameras.However,in real-life scenes,due to the existence of shooting angles,posture change,lighting conditions and other issues,person re-identification becomes a challenging task.The subject of cross-modality person re-identification has also emerged to improve the accuracy in low-light environments.In this paper,we study this topic from two perspectives: network structure and loss function.Compared with the existing methods,the proposed method achieves excellent performance on two public cross-modality person re-identification datasets SYSU-MM01 [1] and Reg DB [2].The specific research contents are as follows:(1)A novel dual-path person re-identification method based on attention mechanism is proposed.Firstly,the attention mechanism is introduced into the visible branch to obtain the feature information of visible image in spatial dimension and channel dimension to achieve infrared feature information matching.Then,the extracted features are horizontally sliced into p parts and mapped to the common space,and the batch normalization neck network layer module was introduced into the cross-modality person re-identification algorithm to reduce the difference of modality feature information and accelerate the convergence speed.Finally,on the basis of hetero-center loss function and cross entropy loss function,the pedestrian identity loss function under cross modality is introduced to improve the accuracy of pedestrian recognition.(2)A novel cross-modality person re-identification method based on dual modality distance constraint is proposed.Firstly,the pedestrian features in different modalities are extracted and the features are cut horizontally.Then,different batch combinations are formed according to the features after cutting.Next,the whole constraint is utilized to narrow the discrepancy among different modalities,and the discrepancy among different classes within the same modality are further enlarged by merging triplet loss and center loss.Finally,modal specific identity loss and cross entropy loss are utilized to improve the performance.(3)In order to compensate for the influence caused by most algorithms focusing only on single coarse-grained or fine-grained features,this paper proposes a cross-modality person re-identification method based on graph convolution.Firstly,the asymmetric full connected network is adopted,and the horizontal cutting in Method 1 is improved to multi-layer feature cutting.Then,the improved graph convolution module is utilized to construct a graph convolution neural network with local and global features as nodes,and the constructed graph convolution neural network is utilized to learn structural features.Finally,Softmax loss function,whole constraint and partial triplet-center loss function proposed in Method 2 are combined to further improve the performance of the algorithm.