Research On Person Re-identification Technology Based On Deep Learning

The task of the person re-identification is to match two pedestrian images captured by two cameras which are not overlapped by the field of vision,which is a key technology in the field of biometrics and security.As for open-set protocol of person re-identification,the ideal person features are expected to have smaller maximal intra-class distance and bigger minimal inter-class distance under a suitably metric.However,this task is extremely challenging due to the large variations of poses,viewpoints,lightings etc.How to extract the pedestrian embeddings of both robustness and discriminative ability is the key step for person re-identification.In this thesis,our work focuses on the person recognition,and specifically is dedicated for the improvement of person re-identification algorithms that based on deep learning methods.The main works of this thesis include the following aspects:(1)In this paper,we use a siamese CNN that combines classification loss with verification loss to construct a joint supervision helps the network to learn more discriminative pedestrian descriptors.To boost the performance,a new feature reweighting layer which in verificationsubnet is designed to explicitly emphasize the importance of each embedding dimension through taking the correlation of across dimensions into consideration,so it gains more freedom to explicitly adjust the scales of the learned embeddings.In addition,a weight constraint is performed on this layer makes the learned embeddings more generation ability.(2)Although the verification loss can improve the discriminability of pedestrian embeddings,it takes image pairs or triplets for training,the number of which grows rapidly as the number of classes grows.It inevitably results in slow convergence and instability.In this paper,we propose a classification network that combines classification loss with center loss to learn more discriminative pedestrian descriptors.The training objective of the classification loss is to separate the learned embeddings,and the center loss pulls the deep features of the same class to their centers.However,the center loss does not consider the distance between different classes.Therefore,this paper penalizes the distances between class centers to improve center loss.With the improved center loss,not only the inter-class features differences are enlarged,but also the intra-class features variations are reduced,which are very essential to person re-identification.(3)In order to verify the practicability of our algorithms,we design a person re-identification system based on matlab and python language respectively.