Person Re-Identification Based On Deep Learning

With the widespread deployment of surveillance cameras,the need for intelligent analysis of surveillance data is rapidly expanding.At the same time,with the rise of deep learning,person re-identification,which is an important recognition technology,has also been rapidly developed.However,there are still some differences between the existing research on person re-identification and actual application scenarios.On the one hand,the existing research is more focused on image-based person re-identification,and there is less progress in the field of video-based person re-identification,while the data in real scenes is mainly video.On the other hand,in the existing datasets,the human body is in most cases complete,but in actual scenes,the human body is often occluded.In this paper,researches are carried out on the spatial-temporal modeling and occlusion issues in person re-identification.The main work and innovations include the following two aspects.Aiming at the problem of spatial-temporal modeling,this paper proposes the relation-guided spatial attention and temporal refinement for video-based person re-identification.In order to cope with the limited information of local operation,this method models and utilizes global information.This paper first designs a relation mod-ule to measure the relationship between the two features,then concatenates all pairwise relationships into a global relationship,and adds them to the spatial and temporal mod-ules.In the spatial domain,the global information is used to learn attention,so that the model pays more attention to the human body area and reduces.the interference of the background;in the temporal domain,the global information is used to update and merge the features to play the complementary role between frames and make the features more robust.Experiments on large-scale video-based person re-identification datasets show that the method in this paper achieves the best performance.Aiming at the problem of occlusion,this paper proposes a method of the partial person re-identification based on region matching.This method uses correlation opera-tions to find the regions corresponding to the partial images in the complete image,and only calculates the similarity of the features of the sub-regions,thereby reducing the influence of other irrelevant regions.The multi-scale matching method uses the feature pyramid to find the optimal matching scale and reduces the influence of the difference in the scale of the two images.Multi-part matching is performed at a finer granularity while maintaining structural information.Related experiments show that this method is extremely competitive and prove the effectiveness of the method.