Research On Methods Of Person Re-identification Based On Deep Learning

Person re-identification(ReID)is to solve the problem of pedestrian image matching across different camera views,which is an important research field of computer vision and intelligent video surveillance.In recent years,researchers have greatly improved the performance of ReID with deep learning and computer vision technology.However,it is still very challenging due to the complex application environment,such as illumination change,complex background variance,low-resolution images,object occlusion,similar clothing of different pedestrians,etc.Based on deep learning technology,this paper proposes several novel methods for ReID from different aspects.The main works and innovation points of this paper include the following parts:(1)For small-scale ReID datasets,combined with multiple instance learning techonology,this paper proposes a deep multiple instance learning model.Because the process of constructing a ReID dataset is very complicated and requires a lot of manpower and material resources,many ReID datasets are usually small.If training a model on these small-scale datasets with deep learning,it would suffer from severe overfitting and degrade the performance of the model.Therefore,this paper proposes a ReID model based on deep multiple instance learning.Firstly,several large local image blocks are extracted from different positions of each person image,and then every image is represented as a multiple instance bag with the guidance of multiple instance learning theory.Then several multiple instance subspaces are formed according to different extraction positions.Moreover,a multiscale convolutional neural network with two parallel convolutional layers in the first layer is designed as the subnet of the siamemse network to reduce the loss of underlying information.Then,an instance-level deep feature embedding model is learned for each subspace by optimizing the contrastive loss function.In the test phase,the average Chamfer distances are computed to obtain the dissimilarities between bags(i.e.person images)for ReID.Comprehensive evaluations on three benchmark datasets(including VIPeR,ETHZ,and CUHK01 datasets)well demonstrate the robustness of our proposed ReID framework on small datasets.(2)For improving the performance of ReID on large-scale datasets,we propose a symmetric triplet loss function and then a hybrid ReID model.Because the image identification based model can make full use of the information of train images.However,because its purpose is to learn the classification boundary,it cannot well maintain the ordering relationships between images and then lead to large inter-class variations.On the other hand,the metric learning based model can maintain the ordering relationships between images,but usually result in an insufficient compactness.Moreover,because it is impossible to sample all triples,it cannot make full use of the training set information.Therefore,this paper proposes a hybrid ReID model beased on image identification and metric learning.Through the joint optimization of Softmax loss and symmetrical triple loss function,the two tasks can cooperate and complement each other to give full play to the advantages of the two methods for achieving a more discriminant and compact feature embedding space.Extensive experiments over several popular benchmarks achieve state-of-the-art results,which well demonstrate the effectiveness of the proposed method for ReID.(3)For ReID in complex scenes,a novel loss function named mini-cluster loss is proposed.The triplet loss and its variants are commonly used in many deep ReID frameworks.However,that those loss functions simply focus on relations of data points may lead to a relatively large intra-class variance and then a weak generalization capacity on the test set.In this paper,we proposes a novel loss function named mini-cluster loss,which regards images belonging to the same identity as a mini-cluster and treats them as a whole during the training.For each mini-cluster in a batch,we define the largest distance between points in a mini-cluster as its inner divergence and the shortest distance with outer points as its outer divergence.By constraining the outer divergence larger than the inner divergence,the framework based on the mini-cluster loss achieves the more compact mini-clusters while keeping the diversity distribution of the classes.The experimental results on two large-scale datasets(Market-1501,DukeMTMC-reID)clearly demonstrate the proposed ReID model based on mini-cluster loss has a higher accuray and better generalization ability.