Research On Key Problems Of Multi-Camera Person Re-identification

The person re-identification is to match the pedestrian targets in the different video cameras in the non overlapping multi camera monitoring system.During the process of the person re-identification,there are such problems as illumination change,visual angle change,pedestrian posture change and even the influence of bad weather,which makes person re-identification face great challenges.This paper makes relevant research on these problems.(1)There are differences of visual angle,light and scale among pedestrians captured by different cameras.The method of person re-identification based on support sample indirect matching is proposed.The algorithm first respectively extracted support samples under different cameras.When matching pedestrian image from two different cameras,based on Euclidean distance and using support samples,the pedestrian category of the pedestrian picture under the corresponding camera is obtained.This avoids the direct matching of pedestrian pictures under different cameras,and fundamentally bypasses the influence of different cameras' view angle,illumination and scale difference.(2)In the process of person re-identification,there are factors such as light,visual angle and other factors.There are more difficult challenges in the real environment.For example,in the fog and haze weather,a large number of pedestrian details are lost,which brings great challenges to pedestrian matching,which is one of the most serious problems in person re-identification.In view of the influence of haze and fog and other bad weather,the method of person re-identification based on dark channel and distance learning is adopted.The method of dark channel prior knowledge is used to remove the fog,and then the local maximum feature and distance learning algorithm is used to reidentify the pedestrians.(3)With the increasing application of deep learning in the field of computer vision and pattern recognition,many researchers have applied deep learning to person re-identification.Deep learning they used is mainly used to extract features that compared to the traditional method are more complex and effective,but these features are based on low level features,such as color and texture.Due to the effects of light,from the perspective of factors such as the different pedestrian between these low-level features are very similar,at this time it is difficult to distinguish.Through the study,it is found that the characteristics of pedestrians can effectively distinguish these similar pedestrians.This paper focuses on the method of combining attribute features with low dimensional features extracted by CNN to identify pedestrians.Using Coupled Clusters Loss training method combining two features of training network,and mapping combined feature to Euclidean space,and using L2 distance can calculate distance between arbitrary pedestrians,so as to judge whether the same pedestrians,which then form end-to-end person re-identification network framework based on deep learning.