| With the gradual improvement of people’s safety awareness,intelligent monitoring equipment has spread to various places.As one of the research directions of video surveillance,person re-identification technology has attracted extensive attention of scholars from all walks of life.Person re-identification refers to using machine learning or deep learning methods to re-identify pedestrians from other monitoring devices given pedestrian images taken by one monitoring device in multiple camera monitoring systems without cross-view.The development of convolutional neural networks has greatly improved the recognition performance of person re-identification networks,but there are still many difficulties,such as light changes,different resolutions and occlusions of pedestrian images,which will affect the network recognition rate.Therefore,focusing on the key problem of "how to improve the recognition effect of person re-identification network",this paper explores how to extract more fine-grained image features and how to better optimize pedestrian feature vectors,and then proposes different person re-identification methods,as follows:(1)A person re-identification algorithm based on softpool and coordinate attention is proposed.First,softpool strategy is introduced in the shallow feature extraction stage for feature map downsampling,which uses an exponential weighting method to retain the basic features of the input pedestrian images,while at the same time amplifying and enhancing the salient features.Second,coordinate attention module is introduced in the deep feature extraction stage,which divides the channel attention into two different one-dimensional feature vector encodings to aggregate features,one retains the location information of the spatial direction,and the other obtains the long-term dependencies of the spatial direction,and finally use aggregated features to enhance the representation of salient information.In order to verify the effectiveness of the method,experiments were carried out on three mainstream datasets,Market-1501,DukeMTMC-re ID,and MSMT17.The experimental results show that the performance of the proposed model has been greatly improved to a certain extent.(2)A person re-identification algorithm based on connected attention is proposed.The algorithm uses the connection attention mechanism in the feature extraction stage.The connected attention mechanism collects information from the previous attention block through the current attention block and passes it to the next attention block,so that the information between the attention blocks can cooperate with each other,thereby improving the learning ability of the attention module.In addition,in the data preprocessing stage,the automatic enhancement strategy is used to expand the data set,which makes the model more robust;in the loss optimization stage,the triple loss of hard sample mining and circle loss are adopted,and finally the model obtained good results.The experimental results further demonstrate the effectiveness of this method on the Market-1501,DukeMTMC-re ID,MSMT17 datasets.(3)A person re-identification algorithm based on an improved triplet loss is proposed.The algorithm uses a feature extractor combined with an attention mechanism and a residual network to extract pedestrian image features,then performs batch normalization on the extracted feature vectors in a common space,and finally uses the angle triple loss function based on cosine metric for optimization.It makes the feature vector separable in the angle space,and can be well combined with the classifier for network training,which finally makes the recognition effect of the network model better.The experimental results show that the algorithm has further improved the recognition accuracy on the Market-1501,DukeMTMC-re ID,MSMT17 datasets. |
