Person Re-identification Based On Deep Regression Analysis

The task of person re-identification is to compare query images with a large number of images in the gallery captured in the non-overlapping video surveillance system,and identify the person in the query images according to the similarity.The basic assumption is that the capturing of person images is finished in a relatively short time,so the clothes and the body shapes do not change much which can be used as evidence of identification.The difficulties of this subject are:(1)There are significant changes in the view of person with the same identity.(2)The resolution of the input images are relatively low,and facial features are difficult to apply.(3)Compared with the training set,the test set is in a fully open set recognition state in person re-identification,so the traditional deep classification models cannot make the person re-identification task end-to-end.(4)There are large changes among person images in lighting,pose,viewing angle,image scale,image resolution,camera settings,and background.We research the above-mentioned challenging problems and propose an end-to-end person re-identification method,which can partly resolve these problems.The main works and contributions of the thesis summarize as following:(1)We take the person re-identification model as a regression model,which makes the person re-identification process end-to-end,and introduce the Deep Regression Neural Network integrating Adaptive Multi-Attribute Fusion method(DRNN-AMAF).The problem that the deep classification model,using the softmax layer as the output layer,cannot handle the open set recognition task,can be solved in this way.First,a shared weight Siamese network is used to extract the differences between the two input images.The differences of these two images are flow into the multi-attribute branch,where each attribute corresponds to each branch of the deep regression neural network,to obtain attribute similarity probability regression.Then,the attribute fusion weight learned by the proposed Bayesian inference based multi-attribute fusion and optimization theory,is used to fuse the probability of each attribute into the final probability that the identities of two person are the same.In order to solve the problem of multi-attribute fusion weights tending to one-hot during optimization,a smoothing-fusion-weight cross-entropy loss function have proposed in this thesis.Applying the loss function during optimization can effectively avoid this problem.(2)In order to solve the problem that the hard-labels provided by the public datasets are not suitable as the supervision information for the training of the regression models,this thesis proposed the Bayesian inference based multi-attribute fusion and optimization theory,which integrated into the deep learning end-to-end training process.This method can adaptively fuse multiple hard-label into a soft-label,and make the attribute label suitable for supervising information during the training process of the regression model,which makes the deep regression model more easily to converge.(3)Compared with classification tasks,person re-identification task requires models to have stronger generalization,due to the open set state.The decision of Convolution Neural Network Based Model(CNNBM)highly depends on the combination of the patterns from the higher layers learned from the training data.Therefore,the diversity of pattern combinations determines the generalization of CNNBM.A Gradient Based Erasing Data Augmentation(GBE)method,which can avoid the model relying on a few specific pattern combinations and reduce over-fitting risks to enhance the generalization of CNNBM,proposed in this thesis.For the proposed method,a large number of ablation experiments and contrast experiments are performed on the Market-1501 and Duke MTMC-reID datasets.The results of the ablation experiments show that,the proposed method has positive impacts on the baseline model and improves the accuracy.The contrast experiments show that the proposed method keeps the accuracy of the model while keeping the model with less parameters and the simple structure model.