Person Re-identification Neural Network Model With Local Feature Fusion

With the deployment of a large number of monitoring equipment in modern society,the manual inspection of monitoring images seems inadequate.The research in the field of intelligent monitoring is attracting more and more attention.Person Re-identification is the core part of intelligent monitoring.The main research content is to find out the designated pedestrian target from the monitoring screen without overlapping field of view,which is used to depict the moving track of the target between monitoring points.The study of Person Re-identification can be regarded as a classification problem similar to Pattern Recognition,which is completed by extracting pedestrian features and feature-metric.Many researches on Person Re-identification focus on the feature extraction of pedestrians.Due to the complexity of pedestrian images,it is often difficult to define a good feature descriptor by traditional methods in the field of computer vision.In recent years,the emerging deep neural network model,although it has a certain degree of pedestrian weight recognition ability,but it is often the learning the image of the global pixel without distinction,unable to distinguish the key areas of the image.Considering the pedestrian images often half pixel belongs to the background image,these pixels are independent of pedestrian targets,affects the quality of the characteristics,so this paper hopes to enhance the impact of pedestrian targets on features and weaken noise interference,so as to improve the quality of pedestrian feature descriptors and obtain a higher performance model based on the deep network model.In this paper,a feature fusion network is proposed based on the idea of representation learning.The local and global features of pedestrians are fused into a unified feature descriptor to identify and classify pedestrians.The network model is composed of three parts: pedestrian local feature extraction part,global feature extraction part and feature fusion part.The local feature extraction module uses an improved LOMO method to describe local features.This method limits the scope of LOMO feature extraction to the main area of pedestrian target and eliminates the interference of background pixel on features as far as possible to strengthen the influence of foreground pixel.The global feature extraction module selects the depth residual network model which has been outstanding in recent years to obtain the global feature vector of pedestrians.The feature fusion module integrates local and global feature vectors of pedestrians through training neurons,and completes the recognition and classification of pedestrian images.Finally,the performance of the feature fusion network is proved to be better than the traditional feature extraction method and the single deep neural network model through the comparative experiment on the mainstream Person Re-identification data set.Then,this paper further improves and optimizes the feature fusion network based on the metric learning method.Based on the feature fusion network,a triplet metric model is proposed to fuse the local features of pedestrians.The feature vectors of triplet are calculated through the feature fusion network,and the feature distance is measured.In order to ensure the training efficiency of the model,the model selects samples from the data set to form the triplet of input by means of hard sample mining.Through comparative experiments,the triplet metric model further improves the classification accuracy of the feature fusion network and achieves the purpose of optimizing the model.Meanwhile,it also proves that the metric learning method can achieve better accuracy compared with the representational learning method in large data sets.