Person Re-identification In Complicated Conditions

In China,the government has invested a lot of money to construct video surveil-lance networks in the city.The way that the police investigate criminal cases have taken place due to the popularity of video surveillance system.As we know,video investiga-tion technologies have achieved a fast development.But,the effect is not equal to the benefit.During the process of the practical video investigation,investigators should watch a huge amounts of surveillance video around the area of the crime place,before and after the crime time,and search for views of the same target,track the target among multiple camera videos,so that find out and trace the suspect.In the past,the video investigation work are done by manual,which costs a large number of human resources and time.To this end,the efficiency requirement of the video investigation pushes the development of person re-identification.Person re-identification is the technology to judge whether a pedestrian in one camera has ever appeared in the other camera.Gen-erally,computer vision and machine learning methods are always used.This technology helps investigators search and track suspects,improve the cases-solving rate,and is of great significance.Recently,person re-identification is becoming a research hotspot,and has obtained very high accuracy under given simulation conditions.However,the conditions become complicated,the effectiveness of person re-identification would be dropped significantly,which cannot fulfill the needs of video investigation applications.The complicated conditions in person re-identification consist of complicated characteristic,complicated distribution,complicated structure and complicated presentation.There still exists several technological difficulties to be overcome under these four complicated conditions.(1)Under the influence of environment and imaging,the resolutions of person images are always different and low in practical video investigation.Sample characteristic becomes complicated.Traditional person re-identification methods based on a uniform resolution will performance bad or become invalid.(2)During practical video investigation,the trained model is always not comprehensive and with low generalization ability,because training samples are limited and insufficient,training and testing classes are irrelevant.Sample distribution becomes complicated.The distance metric function will produce distance measurement deviation on un-trained persons.(3)In general image retrieval researches,the interactive ranking optimization models always rely on global similarity relations.However,in practical video investigation,positive samples and global similar samples are extremely sparse.Sample structure becomes complicated.This will lead the global information based single ranking optimization model invalid.(4)In addition to the feature representation,investigators can mark attributes to represent persons as well.However,experiences and focuses of investigators are different to each other,which make the crowd-sourcing semantic attributes redundant,conflict and fuzzy.Sample representation becomes complicated.Hence,the person representation with attributes are inaccurate.To this end,this paper researches person re-identification in complicated conditions,mainly focusing on multi-dimensional comprehensive representation,data-driven adaptive measurement,combinatorial multi-model optimization,multi-modality fusion representation.The main contributions are listed as follows:(1)Multi-dimensional comprehensive representation based person re-identification method.In the face of complicated characteristic,the performance of the uniform resolution re-identification model will decrease when facing multiple resolution scales.To resolve this problem,we investigate the phenomenon of the changing of distance as the scale varies,and raise the motivation that representing an image pair using multi-scale.To this end,we propose a multi-dimensional comprehensive representation based person re-identification method,which transfers the single scale classification to multi-scale united classification.In this way,the method promotes the ability of multi-resolution re-identification.The experiments show that the distance metric based on a uniform resolution,such as KISSME,will get a bad performance under the multiple resolutions situation,and also show that the proposed method performs well in this first-raised problem,which obtains an obvious promotion on the SALR-VIPeR,SALR-PRID,and CAVIAR datasets.(2)Cross-view consistency excavation based metric adaption method.In the face of complicated distribution,because training samples are limited and insufficient,and training and testing classes are irrelevant,the distance metric function will calculate the distance with a shift.To resolve this problem,we investigate the cross-view consistencies of the same persons,and raise the motivation that making an adaption to each image pair.To this end,we propose to learn different cross-view factors for different image pairs,and a data-driven metric adaption method,which transfers the uniform metric to a data-driven adaptive metric.In this way,the method enhances the ability of data-driven adaptive re-identification.The experiments show that the pro-posed method will promote traditional distance metric methods,such as Mahalanobis distance,LMNN,and KISSME.Comparisons on the public VIPeR,CUHK,and PRID datasets,the promotions of CMC are respectively 3.5-15.2%,2.8-15.1%,and 1.6-9.1%.(3)Region-feedback based combinatorial optimization method.In the face of complicated structure,positive samples and global similar samples are extremely sparse among the ranking results,which makes the single ranking op-timization model based on global information invalid.To resolve this problem,we investigate the sample region structure of ranking results,and raise the motivation that optimization with region-based feedback.To this end,we propose a region-feedback based combinatorial optimization method,which transfers the global interaction to local interaction.In this way,the method increases the ability of combinatorial opti-mization.The experiments show that the proposed method outperforms the automatic re-ranking methods and interactive re-ranking based on global information,and also show that the CMC results will obtain big promotions after each iteration,those are respectively 15-23%and 19-31%on the the VIPeR dataset,and 9-29%and 11-34%on the CUHK dataset.(4)Multi-modality consistency constraint semantic fusion method.In the face of complicated representation,the marked crowdsourcing semantic at-tributes are always redundant,conflict and fuzzy,due to the differences among the ex-periences and focuses of investigators.To resolve this problem,we investigate the key obstacle of the marked semantic attributes,and raise the motivation that fusing the marked and general attributes together.To this end,we mainly propose the semantic fusion method based on cross-modality consistency constraint,and dominance-salience matching model for semantic attribute vectors,which transfers the low level feature representation to high level attribute representation.In this way,the method increas-es the ability of accurate semantic representation.The experiments show that each step of the proposed method is effective,and also show that combined with traditional feature-based method,the proposed method gets the highest accuracy,which is 67.78%at rank-1 on the VIPeR dataset.In summary,this paper overcome the bottlenecks of video investigation by in-vestigating person re-identification in complicated conditions.This paper completes the researches on multi-dimensional comprehensive representation,data-driven adap-tive measurement,combinatorial multi-model optimization,and multi-modality fusion representation.It provides a new way to the person re-identification in practical video investigation on the theory and key technologies.