Joint Metric Learning And Multi-Granularity Feature Extraction Pedestrian Re-identification Research

In recent years,the rapid development of national economy and technological innovation has promoted public safety issues to become one of the hot topics of people's attention.Public places are filled with a large number of security monitoring equipment,resulting in a massive amount of surveillance video.If you continue to use manual investigation methods to obtain valuable information from it,the information may lose its timeliness.Therefore,pedestrian re-identification technology came into being.Pedestrian re-identification aims to give a pedestrian an image,and search the pedestrian across the camera based on the appearance information of the pedestrian in the image,such as wear,posture,and hairstyle.With the rapid development of deep learning,it is applied to the pedestrian re-identification network to improve the performance of the algorithm.However,in real scenes,pedestrian images captured from surveillance equipment are susceptible to factors such as camera angle of view,light intensity,object occlusion,and pedestrian posture distortions,resulting in uneven pedestrian image quality and affecting recognition accuracy.The key to pedestrian reidentification technology is how to combine measurement learning and feature extraction to form an end-to-end re-identification system.Therefore,this paper conducts research from two aspects of joint feature extraction and metric learning.First,a new metric learning is proposed to mutually exclusive features of different categories,and features of the same category are aggregated.Secondly,deep learning is used to design a reasonable and effective network to extract pedestrian strength.The powerful appearance information,and finally the combination of the two,effectively improves the system performance,the main research content is as follows(1)In view of the traditional Softmax loss is difficult to distinguish the difference between the same type of pedestrian data samples,this paper firstly makes improvements based on the Softmax loss,regularizes the weight vector and the feature vector respectively,and introduces the boundary cosine parameter.Secondly,based on the residual network Res Net50 design boundary of Cosine pedestrians re-identification Network(Margin Cos Re ID Network,MCN),learn from the input image pedestrian has difference between global characterization information,so as to realize the end-to-end training the boundary of the modified Cosine Softmax Loss(Margin Cosine Softmax Loss,MCSL)compared with the traditional Loss,not only can make the same category between pedestrian data sample draw close to each other,can also increase the differences between different categories of data samples.(2)In the proposed boundary cosine pedestrian re-identification network,in complex scenes,the similarity of pedestrian posture or pedestrian clothing may cause the extracted features to not have deep-level information,resulting in poor performance of the pedestrian re-identification system.Therefore,in order to propose a Multi-Scale Multi-Granularity Fusion Network(MSMG-Re ID)based on multi-granularity deep feature fusion,this paper uses the residual network Res Net50 as the basic skeleton,and the design includes the global coarse-grained fusion learning branch and local Coarsegrained fusion learning branch and local attention fine-grained fusion learning branch.Through the parameter refinement of the fourth residual of the Res Net50 backbone network,the global coarse-grained fusion learning branch captures the overall approximate attention of the pedestrian from the input image,and learns the multi-level coarse-grained information of the pedestrian;secondly,the local coarse-grained fusion learning branch starts from Extract local features from different regions to learn the coarse-grained and deep-level local features of pedestrians;finally,the local finegrained fusion learning branch introduces the Convolution Block Attention Module(CBAM)to eliminate background interference and force the network to pay more attention to the local features of pedestrians.At the same time,the boundary cosine Softmax loss and the triple loss of difficult negative sample mining are used to jointly train the network,so as to learn more distinguishing pedestrian characteristics.The paper has 29 pictures,14 tables,and 104 references.