Research On Metric Learning Based Methods For Person Re-identification

Person re-identification is a hot topic in the field of computer vision in recent years.The task of person re-identification is to match and identify the pedestrians in different time and place in a non-overlapping monitoring network,so as to achieve the search and tracking of specific targets.A robust and accurate pedestrian re identification system has important practical value for the application of criminal investigation,suspect tracking,search of lost pedestrians in public places and other fields.However,in the process of pedestrian image crossing over the monitoring scene,due to the changes of lighting,shooting angle,pedestrian gait and background,the apparent characteristics of pedestrian image are very unstable,there is the problem of inaccuracy,which makes person re-identification very difficult.Therefore,a large number of scholars at home and abroad are attracted to study this task.In this paper,the main reason that location affects the recognition accuracy of person re-identification problem is due to the small sample problem and the inaccuracy of image features.A multi domain joint depth network model is used to extract the feature expression model,and a new measurement layer is added based on the depth learning algorithm to measure the pedestrian image more accurately.The main work of this paper is as follows:Firstly,the two key contents of person re-identification task,feature extraction and distance measurement,and the existing algorithms are described and summarized.Through programming simulation,the effects of different methods are compared and analyzed.At present,the recognition accuracy of person re-identification problem has reached a bottleneck.Through research and analysis,it is found that the difficulty of person re-identification task lies in the instability of the features caused by the inaccuracy,the complexity of the appearance transformation mode of pedestrian crossing the monitoring scene,and the great difference between the training data and the test data in the metric subspace.Then,the multi-domain joint deep learning network is introduced to extract the apparent image features of pedestrians by combining multiple data sets,and the neuron nodes are selected by the dropout operation method guided by the domain.The feature extraction network model is more stable and has stronger expression ability.In this paper,based on the deep network structure,we add a feedback metric layer,use the idea of migration metric learning,combine with the method of metric learning,optimize the person re-identification model,and design a more effective deep learning person re-identification method.Finally,based on the GUI system design function of Matlab,this paper simulates the person re-identification application system design,including the introduction of Matlab GUI design function,the design requirements of person re-identification simulation system,and the design scheme,and gives the simulation effect through the actual test.The test results show that the 'deep learning+metric learning' person re-identification model in this paper is better than the comparison method.Compared with Metric Learning by Accelerated Proximal Gradient(MLAPG)algorithm,the accuracy of rank-1 is improved by 16.46%on VIPeR dataset.