Research On Person Re-identification Based On Multi-feature Fusion

In recent years,with the improvement of computing power of computers and the storage capacity of storage devices,as well as the demand for security in various industries,more and more surveillance cameras are arranged in different scenes.The manual screening of massive video data is difficult and cannot meet the demand.Intelligent video surveillance is particularly important.Person re-identification is an important research branch in the field of intelligent video surveillance and is a hot and difficult point in the field of computer vision.Person re-identification means recognizing the same target pedestrian in different camera systems without overlapping.Due to the different attributes of the camera,the shooting angle and the shooting time lead to the change of illumination,angle of view,pose and scene of the pedestrian target image,which increases the difficulty of person re-identification.The recognition rate is a measure of person re-identification algorithm.This paper aims at the characteristics of pedestrian target image and to improve the pedestrian recognition rate,a person re-identification algorithm based on Laplacian kernel canonical correlation analysis of multi-feature fusion is proposed.Person re-identification mainly focuses on feature extraction and matching metrics.Efficient feature descriptors help to improve the recognition rate.In the feature extraction stage,this paper combines local maximum occurrence features,covariance descriptors based on bio-inspired features,weighted overlapping stripes histograms and improved local feature sets.In order to overcome the illumination,pose,and the change of view angle,and form a feature descriptor with robustness and validity.Before the matching measure,most of the existing literature will perform principal component analysis to reduce the dimensionality of the extracted features,but does not consider the relationship between different instance features.In this paper,a novel feature dimensionality reduction method is used:Cross-view Quadratic Discriminant Analysis can obtain the feature subspace with strong discriminative power;in order to further find the same category,the Laplacian kernel function is used to map the feature to another non-linear space.At the same time,a canonical correlation analysis is used to learn a common subspace.Finally,the similarity measure is completed by the Euclidean distance.The proposed method has been tested and simulated on the challenging VIPeR,PRID450s,and CUHK01 public datasets.The experimental results show that the proposed method is competitive with the existing advanced algorithms:the rank-1 in VIPeR,PRID450s and CUHK01 datasets reach 51.23%,70.04%and 55.82%respectively,it is proved that the proposed method is superior to many typical advanced person re-identification algorithms,and is feasible and effective.