Person Re-Identification Based On Multiple Cameras Without Overlapping Area

Person re-identification under multiple cameras without overlapping area is the process of searching for people interested in a camera in other non-overlapping cameras,and it is an important part of intelligent video surveillance.Person re-identification plays an important role in graphic search,criminal investigation and other aspects.Because of the change of person posture,the low resolution of picture,the change of environmental,the accuracy of person re-identification is low.Therefore,it is of great significance to improve the accuracy of person re-identification.In order to improve the accuracy,for the problems that the neural network framework is difficult to design and difficult to converge during training,this paper designs the transfer learning based on the Alexnet network model,improves the existing model,and combines it with the XQDA(Cross-view Quadratic Discriminant Analysis)algorithm.Aiming at the traditional method,a new background suppression method is designed to reduce the influence of background and improve accuracy.The main research work is as follows:(1)Research methods for person re-identification at home and abroad: method based on feature representation,method based on distance measurement,and method based on deep learning.We compare the advantages and disadvantages of the three methods and find the direction of improvement.(2)For the problems that the neural network model structure is difficult to design and difficult to converge,this article is based on the Alexnet model for transfer learning.By improving the last layer of the Alexnet model,it can adjust the parameters adaptively according to the task requirements of person re-identification,so as to realize the migration learning,at the same time the network is easy to converge.By extracting the deep feature representations of the second fully connected layer of the neural network model and using the XQDA algorithm for similarity measurement,the accuracy is further improved.Experiments are performed on the CUHK03 dataset and the DukeMTMC-reid dataset.The CMC-1 is as high as 56.2% and 53.53%,respectively.(3)The accuracy of person re-identification is not high with the use of deep learning methods for small sample data.This paper has improved the traditional method and designed a background suppression method.This method combines the HOG features with random ferns to calculate random ferns features in the HOG-domain.The K-means clustering algorithm was used to classify the random fern features in the HOG-domain,and the background classes are filtered and removed to achieve background suppression.Then the color histogram feature and SIFT feature are extracted and the XQDA algorithm is used to measure the similarity between different features.Experiments are performed on data sets such as viper and CUHK01.Compared with the method without background suppression,the method of this paper improves CMC-1 by 8 percentage points and 6 percentage points respectively on the two data sets.And compared with the method of Chapter 3,CMC-1 increased by 2.5 times and 2 times respectively.