Research On Several Algorithms Of Person Re-identification Based On Metric Learning

In video surveillance,when a pedestrian is captured by a camera located in a public place,the process of using the existing camera network to discover the location of the target pedestrian next time is called person re-identification.It has important application significance in criminal investigation,security,and other public activities.However,due to non-overlapping cameras,large changes in pedestrian angles,complex backgrounds,and occlusion,person re-identification is challenging and has become a hot spot in computer vision.In recent years,many scholars have proposed many algorithms.Based on the existing algorithms,this paper further studied the person re-identification algorithm based on metric learning.The main contents include:1.First,the commonly used methods of person reidentification are researched and compared.Aiming at the over-fitting problem caused by the imbalance of class in the current person reidentification problem,the method of clustering centralization relieves the influence of the class imbalance so as to improve the recognition accuracy.In addition,this paper uses principal component analysis for pedestrian features to reduce the feature dimension to an acceptable dimension to improve the training speed,and reduces the impact of noise on the training process by reducing the dimension.Experimental results show that this method can effectively improve the training speed and accuracy.2.A new person reidentification method based on iterative projection vector learning was presented,which is based on the problem of higher time complexity and storage complexity of the commonly used training Mahalanobis matrix based on metric learning.An orthogonal updating strategy for feature distribution was proposed in this method,which utilizes this tactics to be able to guarantee that the projection vector that is updated every time is orthogonal.Moreover,the perturbation term is applied to the updating strategy to ensure that the projection vectors are approximately orthogonal,so as to preserve the relation between the projection vectors and make the projection more practical.Therefore,a small part of the projection vectors can be used to form a low-rank projection matrix,so that the features after projection are more discriminative.Experiments showed that the proposed algorithm has relatively low requirements on the features and can obtain better experimental results than other algorithms.3.According to the temporal and spatial distribution characteristics of the video sequence dataset features,based on the KL divergence(Kullback-Leibler divergence)discriminant model,assuming that all pedestrian features share the covariance matrix,the characteristic divergence between different pedestrians is maximized,So that different pedestrian features have better separability,and for the same pedestrian constraints between the covariance matrix and the initial covariance matrix as close as possible,so as to retain the relations between same pedestrians,making the model has better robustness.The experimental results showed that the algorithm can well retain relations of the same pedestrian and increase the difference between different pedestrians,and it can obtain high recognition accuracy in the sequence dataset.4.At present,metric learning based person reidentification needs a large number of labels,which is not conducive to the model promotion.What's more,recogonition rate of supervised learning based methods is higher than that of unsupervised learning based methods.To address these issues,combining metric learning and dictionary learning,semi-supervised learning dictionary rectification method for person reidentification has been proposed.First,projection matrix learning term is built by few labeled samples,and difference term is built by projected feature vectors and sparse coefficients of dictionary learning,which can rectify the feature dictionary and make it more discriminiate.Then,using projection matrix to project the unlabeled features,so the samples are re-marked,which transforms the semi-supervised learning problem to the supervised learning problem.Experiments results showed that the algorithm has higher recogonition rate than other unsupervised learning based methods.