Improved Reranking Approach for Person Re-identification System

Abstract

AbstractSafety is ensured in all the public places with the help of surveillance cameras. It can be used to find any person involved in crime or lost person. Human operators do this job by enabling correspondence between images of same person captured across diffrent cameras. Automation of the same is called as person re-identification (PRID). One of the major challenges of the PRID process is to determine the representation of images which discriminates the person identity irrespective of the view-points, poses, illumination variations, and occlusions. With this motivation, a new deep learning-based inception network (DLIN) with ranking is proposed for person re-identification (PRID) in surveillance videos. The DLIN technique mainly uses Adadelta with Inception-v4 model as feature extraction technique in which the hyperparameters of the Inception-v4 model are optimally tuned by the use of Adadelta technique. The features are extracted from the probe and gallery images. Euclidean distance-based similarity measurement and expanded neighborhood distance reranking (ENDRR) is employed for determining the similarity and ordering the output of the PRID process along with Mahalanobis distance. A wide range of simu-lations are carried out using CUHK01 benchmark dataset and the experimental outcomes ensured the goodness of the DLIN technique over the other existing techniques.KeywordsPerson reidentificationDeep learningSimilarity measurementSurveillanceReranking process