Research On Person Re-identification Method Based On Metric Learning

Person re-identification is the task of finding a person of interest across a network of cameras.Due to its wide applications in forensic search,missing person finding and cross-camera tracking,person re-identification has drawn a lot of attentions in recent years.In person re-identification,the appearance of individuals is the mainly exploited information,since more reliable biometrics such as face,gait or fingerprint are not always available in typical surveillance scenarios.However,the appearance of individuals may undergo drastic changes because of viewpoint changes of different cameras,partial occlusion of body parts,variations in illumination conditions and background clutters.In this dissertation,we investigate the problem of how to learn a robust metric to reflect the difference between person identities.Firstly,based on the fact that people care more about the correctness of images at the top of the ranking list,we propose a top-heavy ranking loss to measure the difference between current ranking list and the ground truth ranking list.The top-heavy ranking loss assigns large weights to top positions of the ranking list,which is more suitable for person re-identification.Moreover,we introduce an explicit nonlinear transformation function for the original feature and learn an inner product similarity under the structured output learning framework.Experimental results on several public person re-identification datasets demonstrate the effectiveness of the proposed approach.Secondly,traditional metric learning based person re-identification methods only adopt similarity/dissimilarity labels to learn distance functions,but fail to exploit the latent relative similarity relationship between images.In this dissertation,we decompose the similarity between two images into two parts:content similarity,which measures the likeness between image appearance,and context similarity,which reflects the commonness of their neighborhood.We further propose a contextual similarity regularized metric learning method for person re-identification,which motivates similar pairs to share similar context.Experimental results on several public person re-identification datasets show that the proposed method can promote the performance of person re-identification effectively.Thirdly,we summarized the strategies of conventional metric learning methods and propose three characteristics of the target sample distribution,namely intra-class compactness,inter-class separability and high generalization ability.We further pointed out that current metric learning methods can be further improved by minimizing intra-class distances to enhance intra-class compactness,as well as maximizing the minimum inter-class distance to promote the generalization ability.We transform the above motivation into equidistance constraints,and propose the equidistance constrained metric learning method for person re-identification,which force intra-class distances to be 0,and inter-class distances to be a constant value.The square loss is adopted to form the objective function,which is solved by the projected gradient descent method.This simple method is proved to be very effective in person re-identification task.At last,to learn feature representations and similarity measures jointly,we propose the deep metric learning method for person re-identification.The proposed architecture is composed of a feature extraction network which combines four convolutional neural networks(CNN),each of which embeds images from different scale or different body part,and a similarity layer which calculates the cosine similarity between images.The whole network is trained in an end-to-end fashion with an adaptive margin ranking loss.By introducing an adaptive margin parameter in the ranking loss function,it can assign larger margins to harder negative samples,which can be interpreted as an implementation of automatic hard negative mining strategy.