Research On Method Of Person Re-identification Based On Spatial-temporal Information Re-ranking

In recent years,person re-identification has attracted more and more attention because of its importance in the field of public safety.Person re-identification aims to judge whether two images from different non-overlapping cameras contain the same pedestrian.However,because of the so much variance between the different cameras,such as changes in body pose,camera angle and illumination conditions,identifying the same individual across different camera views has not been solved yet.First,most existing person re-identification approaches assume all images are aligned,so they extract the features and calculate the distance directly.However,because of the automatic detection algorithm,these images are not aligned perfectly.And feature alignment is a key problem to improve the performance.Second,most existing methods treat the initial rank list as the optimal result,which is thoughtless.Generally speaking,the initial rank list is sub-optimal and there is potential information to optimize the initial rank list to get a better result.In this dissertation,our work focuses on improvement of person re-identification algorithms that based on deep learning methods for feature alignment and re-ranking methods.The main work of this dissertation is listed as follows:(1)By introducing a pose estimation model to detect key points,segment the local parts of the human body with these key points.Then train the network branches with local parts respectively to extract discriminative local features.Then combine these local features to achieve local features alignment.Finally,combing global features and local features to form more robust and generalized.(2)We propose a novel re-ranking method through exploiting the spatial-temporal information.In contrast to all existed re-ranking methods,the proposed method bring in spatial-temporal information in the re-ranking stage and treat it as auxiliary information to optimize the initial rank list.The proposed approach is full-automatic without any human feedback.All it needs is spatial-temporal information which can be easily stored in the name of images,so it can be implemented to any baseline to improve the performance.(3)For the network and re-ranking method proposed in this dissertation,experiments are conducted on several datasets,which prove the performance of the proposed network and re-ranking method by comparing the results with other existed methods.