Research And Application Of Pedestrian Re-identification

Person re-identification is designed to conduct cross-camera retrieval for a person of interest.With the constant development of video surveillance network,this technology has received more and more attention from the scholars.At present,person re-identification mainly encompasses three key problems including pedestrian detection,which focuses on extracting high-quality static pedestrian images from the video stream;feature expression,in which feature vectors of cross-camera invariance are extracted from the images;metric learning,which is used to evaluate the similarity of feature vectors.The research on person reidentification is launched hereof on the basis of aforesaid key problems.In respect of pedestrian detection,as typical HOG+SVM algorithm has the following disadvantages including failure in consideration of detection speed and quality and redundant pedestrian image,a pedestrian detection and tracking algorithm composed of motion detection,object classification and object tracking has been realized to meet the requirements in the reidentification scene.For feature expression,with regard to the disadvantages of LOMO feature such as poor robustness of background and high dimension,an improved algorithm introducing pedestrian contour information and dimension optimization are proposed.The improved LOMO feature we obtained has lower dimension and higher discrimination than that of original LOMO feature.The feature expression algorithm based on deep learning is researched,and regarding how to take advantage of complementary properties of improved LOMO feature and CNN feature effectively,a feature fusion network is proposed.With the help of feature fusion network,the improved LOMO feature can realize adaptive fusion with CNN feature and reversely optimize expression ability of CNN feature so as to obtain a fusion feature with higher discrimination.In terms of metric learning,some metric learning algorithms of person reidentification field are researched and their matching rate with fusion feature is evaluated,of which the data result indicates that XQDA algorithm is highly matched with the fusion feature.Finally,on the basis of theoretical research and B/S framework,a person re-identification software applicable to surveillance scene is designed.The software realizes video playback and re-identification and further verifies the correctness and validity of relevant algorithms.