Regularized Smoothing And Saliency Weighted Metric Learning For Person Re-identification

With the popularity of video surveillance equipment in recent years,person re-identification has been paid more and more attention This technology has high application value in criminal investigation,security monitoring and some commercial applications.At present,person re-identification has two main directions:feature representation and metric learning.In view of the above two directions,this paper proposes several innovative methods.Aiming at the direction of feature representation,we propose a pre-feature fusion method,a late-feature fusion method,an optimization method and saliency weighted feature representation method.For metric learning methods,we propose a saliency weighted metric learning method to learn a new metric matrix.Regularized smoothing technology is introduced to enhance the robustness of metric matrix,and a metric learning method based on the saliency weighted of regularized smoothing is proposed.The following is a summary of the research work in this paper:(1)Mulki-feature Fusion and Weight Optimization for Person Re-identification:According to the characteristics of different features,the features are combined. The combination is based on the similarity between different features.Combining the two features with smaller similarity can achieve the effect of complemantary advantages and improve the recognition accuracy.Based on this,a pre-feature fusion method is proposed.After feature combination,multple combination features will be generated.Next,we propose a late-feature fusibn method,which combines multple combined features into one.The weight optimization algorithm is proposed to optimize the performance of the fusion similarity funtion.(2)Saliency Weighted Feature Representation and Local-Global Optimal Fusion for Person Re-identification:According to the characteristics of different saliency methods,AC saliency method is used to extract the overall saliency value of pedestrian images,and FT saliency method is used to extract the saliency value of local blocks of pedestrian images.According to this principle,the saliency weighted feature representation(SWLOMO)is proposed. When the SWLOMO method proposed in this chapter extracts pedestrian image features,AC saliency algorithm and FT saliency algorithm are used to extract global saliency and local saliency respectively,so it is time-consuming.In order to improve efficiency,a metric learning method based on saliency weighting is proposed.This method only needs to extract the saliency value of pedestrian pictures in the training stage and learn a matric matrix based on saliency weighted metric learning.In the test phase,the metrics matrix is directly used to match and identify,which saves the time consumed to extract the saliency value of the test image.When learning saliency weighted metrics matrix,the saliency values of two pedestrian samples are averaged arithmetically,and the processed saliency values are integrated into the learning metrics matrix.The purpose of arithmetic average processing of saliency value is to reflect the commonness of the same pedestrian sample.Because the saliency of the same pedestrian sample will have many similarities,the arithmetic average processing of the saliency of the same pedestrian sample can effectively improve the recognition effect of the same pedestrian sample.Finally,the pedestrian image is divided into upper and lower parts,and the SWLOMO method is used to extract the features of the upper and lower parts and the whole picture respectively.The local and global features are fused by the later feature fusion method and the weight optimization algorithm.(3)Regularized Smoothing and Saliency Weighted Metric Learning for Person Re-identification:We propose an improved method for the saliency weighted method in Chapter 3.The original pedestrian metric matrix is introduced into the saliency weighted method.We hope that the original pedestrian metric learning and saliency weighted metric learning can be fused according to appropriate weights,so as to improve the recognition accuracy.Then,we introduce the regularized smoothing technique to modify the covariance matrix involved in the learning metric matrix.Then,we apply the regularized smoothing technique and the improved saliency weighted metric learning method to XQDA metric learning method and LSSL metric learning method,and get the regularized smoothing and saliency weighted XQDA metric learning method and the regularized smoothing and saliency weighted LSSL metric learning method respectively.Finally,we use the weight optimization algorithm of Chapter 2 to fuse several metric learning methods based on regular smoothing and saliency weighting.