Multiple Back Propagation Network And Metric Fusion For Person Re-identification

With the rapid development of computer vision in the field of monitoring,more and more computer researchers have targeted the exploration of person re-identification methods,making this hot spot a relatively large research topic in the field of computer vision.The main function of person re-identification technology is to judge whether the persons recorded at different times,different locations,and different cameras have the same identity.This technology is applied to intelligent security,criminal surveillance,space-time trajectory tracking and business district management.Challenges faced by the research institute of this technology are face occlusion,low-resolution video,shooting angle,lighting and severe fragmentation of human trajectories in a wide range of space,resulting in ineffective use of video data and so on.At present,research scholars focus on the two main research directions of person re-identification technology:person feature representation and metric learning.This thesis focuses on these two research directions and innovates several effective methods to improve person identification rate.In extracting the person feature representation direction,we propose a multiple back propagation network with several classic deep networks as the kernel,and each stage performs back haul to improve the feature fineness.In the direction of metric learning,we propose a method of reordering Euclidean distance based on Euclidean distance,and optimize weighted fusion with several other traditional metric methods to obtain a more robust matrix,thereby improving the person identification rate.The main contributions made in this thesis are summarized as follows:?1?Research on the characteristics of person re-identification based on multiple back-propagation networks:Considering that during the process of deep network model's propagation from shallow to deep,the feature size of person pictures will decrease,which will cause a loss of some picture feature information,reduce measurable data,and a part of recognition accuracy.Based on this phenomenon,we propose a method of multiple back propagation network?MBP?learning.The network framework of the MBP method is extended based on the original ResNet50^[1]and DenseNet121^[2].We take DenseNet121 as an example.Each Dense-conv layer is connected to an MBP sublayer,and each MBP sublayer is Divided into two sub-flows,one of which is connected in a fully-connected manner and used to calculate the Softmax loss,the other sub-flow is connected to a conv layer to reduce the size after passing through the fully-connected layer,and is used to calculate the Tripletloss.The features are concatenated as the final feature of the person picture.According to the experimental results,the features of the multiple back propagation network are significantly improved compared to the features extracted by the original deep frame.?2?Research on person re-identification algorithm based on re-ranking Euclidean distance:Considering that the traditional distance measurement algorithm is simply a linear measurement,the recognition rate is low,and it cannot solve problems such as full-automatic and unsupervised.This brings great trouble to multiple target person classification,so we propose to reorder the Euclidean distance?Re-ranking Euclidean?.First,the person picture is encoded into a single vector through its-reciprocal nearest neighbors,and then its reciprocal features are calculated.Finally,the Euclidean distance is calculated by reordering.In short,it is to define a candidate set and a query set,find the nearest neighbors of the query picture in the candidate set,then calculate the inverse neighbors of the order,and finally bring the original Euclidean distance formula to get the reordered Euclidean distance.Experimental results show that the effect of reordering Euclidean distance is significantly improved compared with traditional distance measurement algorithms.?3?Research on person recognition algorithm based on metric fusion:A single deep person feature may be able to achieve good experiments,but each depth model has its own advantages and disadvantages.Considering this problem,we propose a method for late fusion of person features called metric fusion?MF?.The Euclidean distance is combined with several other distance measurement algorithms to perform weighted fusion with the person features extracted from two different networks to obtain the final metric distance.In addition,we also define an automatic learning weight optimization algorithm that can make the same results after fusion.The distance between the features of an identity person picture is as small as possible,and the distance between the features of different identity person pictures is as large as possible,thereby further improving the recognition rate.Experimental results show that the MF method has a certain improvement over the single-metric algorithm.