Dynamic Metric And Multidimensional Representation Based Person Re-Identification

In recent years,with the rapid development of China's market economy,the uncertainty factors which affect the public security environment is also increasing,the present public security situation is still at a very serious stage.In order to protecting people's life and property safety,and more effectively to crack down on criminal offenses,video detection technology has become an important technology means for the public security department.However,traditional video detection always using artificial browsing way to seek the clues in the surveillance videos,it requires a lot of manpower and time,and it can't meet the demand of "fast detection and fast break" for the criminal investigation.So the retrieval of a specific target(especially human)based on the surveillance videos has become an important problem to be solved in criminal investigation.Under this background,our research topic--person re-identification aiming at recognizing an individual from images observed across non-overlapping cameras,has attracted increasing attentions these years.However,due to the particularity of the practical surveillance scenario,person re-identification problem faces many challenges.For example,first,designing a set of features that are both distinctive and stable is extremely difficult,specially,sunder realistic conditions where the viewing changes can cause significant intra-object appearance variation.Second,the resolution of the pedestrian images is low and vary,which makes the performance reduction with the traditional methods based on high and coincident resolution.According to the above problems,this paper launches research from the "dynamic metric" and "multidimensional representation" two aspects,research results are mainly as follows:(1)Adaptive margin nearest neighbor for person re-identificationFor the problem that the practical surveillance environment is complicated,and it's hard to get robust and effective image representation,this paper adopts metric learning methods,which learn a distance metric or projecting features from different views into a common space to suppress inter-camera variations.However,general Mahalanobis metric learning methods,which treat all the impostors equally,have ignored some important phenomena for person re-identification.To solve this problem,we propose an Adaptive Margin Nearest Neighbor(AMNN)method to push different impostors away with adaptive variable margins.With the benefits of the "dynamic metric",our proposed method AMNN achieves a better performance.(2)Scale distance surface function for person re-identificationFor the problem that the resolution of the pedestrian images is low and vary in the practical surveillance scenarios,we propose a scale distance surface function method adopting a"multidimensional representation" idea to transfer the single image identification into multi-scale joint identification.By the scale distance surface classification model we trained,our SDSF method achieves a better performance under the multi-resolution conditions.(3)Public person re-identification systemBased on the proposed AMNN and SDSF method,this paper designs and implements a public person re-identification system.Our system has a more comprehensive retrieval ways compared with the existing person re-identification systems,so it's more suitable for different video surveillance scenarios.As there are more than one query image for the target pedestrian in practical video surveillance,our system supports multiple query images joint retrieval.And in order to assist investigators screening of right retrieval results,our system supports playing the search results' video clips,which can provide an additional continuous information of pedestrians walking motion to investigators.As above,this paper mainly focuses on person re-identification technology in surveillance video,and proposes solutions for practical application problems.It plays a vital role in improving the algorithm performance,further promoting the development of video surveillance technology,and maintaining public security,cracking down on illegal crime and protecting people's life and property safety.