Person Re-identification In Video Synopsis System

With the increasing awareness of security and the rapid development of digital video technology,video surveillance network plays a pivotal role in harmonious society.Video synopsis system which can be used to concentrate surveillance videos makes information mining easier.However person re-identification in video synopsis system is a challenging due to viewpoint variations,low-resolution and illumination changes.Thus,in this study we make an expansion on car re-identification based on video synopsis technology.The paper makes three key contributions:First of all,we design an unsupervised real-time person re-identification pattern.We use two-level search structures to deal with mass pedestrian images in surveillance videos.On first level,we extract dense features,fuse similarities of multiple features by RankSVM model and get a small range of pedestrian candidates.In second level,we extract consuming and complex local features to match details.Finally,we get the result images for person re-identification.In this paper,we design a fast and distinguishable local feature MLD-VLAD.To evaluate the effectives of our proposed approaches,we carried out our experiments on our dataset.The results demonstrate that MLD-VLAD with color space outperform SIFT.Secondly,we design a supervised real-time person re-identification pattern.We propose distance metric learning to our unsupervised real-time person re-identification pattern.The traditional distance metric learning measure aims at learning a Markov matrix and calculates Euclidean similarity between pedestrians in new feature space.We propose a novel method named Largest Margin Nearest Neighbor Based on Pearson Correlation Distance(p-LMNN)to improve the performance of person re-identification.P-LMNN learns a linear transformation matrix,projects original features onto a low-dimensional subspace and calculates Pearson Correlation similarity between pedestrians.This distance metric learning method can reduce the dimension while ensuring same person closer and different persons farther.Moreover,it overcomes the unrelated issues of Euclidean distance metric.To evaluate the effectives of our proposed approaches,we carried out our experiments on VIPeR dataset.The results demonstrated that our approach can enhance the performance of person re-identification.At last,we design a real-time car re-identification pattern as an expansion of person re-identification pattern.Since features of cars are more rigid than those of person,an unsupervised method can obtain satisfactory results.The real-time car re-identification pattern is suitable for any scene which makes it replicable.