Person Re-identification Based On Convolutional Neural Network

In video surveillance system,person re-identification is a very important task.The goal of pedestrians recognition is to establish the contact of pedestrian pictures in different cameras under different cameras,to identify whether a specific pedestrian appears on a different camera,and to obtain a specific pedestrian path,which is of great practical significance.In practical applications,the gap between the same pedestrians under the different cameras is larger because of the different camera angle scenes,light changes and the change of pedestrian attitude.The pedestrian weight recognition becomes a challenging problem in the field of video surveillance.In traditional methods,pedestrian weight recognition is divided into the method of feature extraction based on artificial design and the measure learning method of learning the distance function by learning the similarity of pedestrians.In recent years,in the field of computer,the advantages of convolution neural network in learning features have been fully demonstrated,and have shown excellent performance in such tasks as image classification,target detection and so on.In this paper,we mainly study the method of person re-identification based on convolution neural network.(1)A method of person re-identification based on Siamese convolution neural network is proposed.The popular convolution neural network is used to extract the distinguishing feature of pedestrians,which is difficult to extract the discriminative characteristics of pedestrians in traditional methods.On the basis of the classical convolution neural network(AlexNet),combining the advantages of Inception-2 and Inception-3,a convolution neural network for pedestrian feature extraction is proposed in this chapter,and the final pedestrian characteristics are used to carry on the pedestrian weight recognition.In view of the shortage of pedestrian data set,the Siamesenetwork structure is used to increase the number of samples by increasing the number of negative samples to solve the over fitting problem caused by the lack of training samples in the network training process.The input of the Siamese network structure is a pedestrian sample pair.By training a similarity model,the similarity between the same pedestrians is close,and the similarity of different pedestrians becomes larger.(2)A method of person re-identification based on DGD(Domain Guide Dropout)convolution neural network is proposed.In the field of person re-identification,the size of a single pedestrian data set is often smaller,and data collection is from similar scenes with the same underlying distribution.The convolution neural network is used to train the single pedestrian data set.The training model needs to consider how to avoid overfitting and not have the universality,and the practical application value is general.The DGD convolution neural network obtains multi domain data sets by mixing multiple data sets and redistributing labels.The multi domain data set is used to train the convolution neural network to prevent the overfitting of the convolution neural network due to the small number of samples.The learning model can carry on the person re-identification in many scenes,and it has the practical application value.After extracting the features of 256 dimension by using the DGD convolution neural network,this chapter presents the relative distance of the sample feature.By calculating the relative distance between the features of each sample,the feature of discriminating the inter class distance between classes is screened in the 256 dimension feature,in order to improve the effect of the pedestrian weight recognition algorithm.In the 256 dimensional features,128 and 192 dimensions are selected in this chapter to test the characteristics of discriminant distance between classes.The algorithm proposed in this paper is experimentation on public image dataset CUHK03 and iLIDS,which proves that the algorithm is effective.