Research On Key Techniques For Person Re-Identification Under Complex Scenarios

Due to the rapid development of society security and development of video surveillance,person re-identification(person re-id)plays an important role in security surveillance,criminal investigation and target recognition applications,and has attracted extensive attention of machine vision and artificial intelligence.A large number of researchers have investigated the person re-identification problem,and achieved interesting results.Most works mainly focus on person re-identification under normal scenarios,i.e.,little change of distance between cameras and pedestrians,little change of illumination and few occlusions for pedestrians.However,the complex factors in real-world scene,e.g.,the changes of weather,pedestrian movement and resolutions of cameras,increase the difficulties of person re-identification task.(i)Due to the limitation of storage capacity and the possible fault of hardware,the captured samples only contain the grayscale information without RGB color information.There exist no public grayscale video datasets for person re-identification task to date.Few works have been presented to improve the recognition by compensating for the color loss of grayscale video.(ii)Due to the low illumination at night,the inadequate exposure leads to less information of samples than that under normal illumination.Although some public datasets contain the changing illumination,there exist no standard lowillumination pedestrian video dataset captured at night.How to reduce the effect of lowillumination on sample is an important task for person re-identification.(iii)In practice,the resolutions of samples are different due to the differences between different cameras,and the distance between cameras and pedestrians.High-resolution samples contain much effective information for person re-identification,while low-resolution samples contain little effective information.Most existing works focus on the matching of different resolution images,but there exist few studies for different resolution videos based person re-identification.(iv)For security application scenes such as criminal investigation,it is difficult to obtain image samples of the target suspect.The descriptions of the suspect's appearance information by eyewitness can help person re-identification.The eyewitness can only provide rough descriptions about the color and style of the suspect's clothing and appearance.The details of color-illustration-style samples drawn by the descriptions are rough.However,it can uncover the relationships between color illustration and real photos.At present,there exist few studies on person re-identification between color illustration and real photos.This thesis analyzes the person re-identification problems under the complex scenarios as the above,and proposes corresponding solutions.In the process,we have achieved some valuable results:(1)To tackle the problem of grayscale and true-color video based person re-identification and reduce the loss of effective information in grayscale video,this work proposes a semicoupled mapping based dictionary learning.Specifically,we learn a within-video projection matrix for each pedestrian video,then learn grayscale dictionary and true-color dictionary respectively.Then we learn a semi-coupled mapping matrix between grayscale and truecolor samples to establish the intrinsic relationship between the heterogeneous samples.To improve the discriminative ability of dictionary pair,we design the fidelity item and regular terms.In addition,we have recorded and released a new grayscale and true-color video person re-id benchmark dataset,which can provide rich data for further researches.(2)To handle the problem of person re-identification under low-illumination scenario,we propose a triplet-based manifold discriminative distance learning.Specifically,we divide each pedestrian's video sequence into several local linear models by geodesic distance,and learn the distance metric.With the multiple constructed local linear models,the learned distance metric can minimize the intra-class distance and maximize the interclass distance.To fill the gap of the low-illumination pedestrian dataset,we have collected and published a new low-illumination pedestrian video dataset.(3)To tackle the low-resolution and high-resolution video based person re-id problem,we propose a semi-coupled mapping based set-to-set distance learning method.Firstly,we divide the video sequence into several subsets by walking cycles to extract features,then learn a semi-coupled mapping between low-resolution and high-resolution features,and finally learn a set-to-set based discriminant metric.To handle the loss of sample with low resolution,we design a semi-coupled mapping term to compensate for the loss of sample effective information.We learn a set-to-set based discriminant metric to minimize the intraclass distance and maximize the inter-class distance.We have collected and published a new low-resolution and high-resolution video pedestrian benchmark dataset under real-world scenarios.(4)To deal with the problem of color-illustration-style image and normal-photo-based person re-identification,we propose a semi-coupled mapping and discriminant dictionary learning.The model can learn the dictionary pairs for color-illustration and normal-photo samples respectively,and then learn the mapping relationship between heterogeneous sample pairs to reduce the loss of color-illustration samples.The color-illustration samples and normal photos are generated in different ways.Therefore,they are heterogeneous.To handle the problem of heterogeneous samples,a semi-coupled mapping term is designed to uncover the intrinsic relationship between heterogeneous sample pairs.The semi-coupled mapping matrix can reduce the difference between heterogeneous samples and improve the matching.To provide a real-world pedestrian database with color-illustration-style image and normal photo,this thesis has collected and drawn a new color-illustration-style image and normal photo pedestrian re-id dataset(CINPID).