The Research On Automatic Annotation Algorithm For Person Re-Identification Via Visual Attention Mechanism

Person re-identification has many practical applications in computer vision and pattern recognition,such as intelligent monitoring,criminal investigation,material evidence tracing,theft analysis and evidence collection,searching for lost children or the elderly and preventing terrorist attacks.However,most of the existing person re-identification researches are conducted in closed-world rather than in open-world,which greatly limits the practical application of person re-identification technology.The open-set person re-identification driven by practical application is to find a person identity to be inquired in the whole frame image.Through the intelligent analysis function of video surveillance,the useful intelligence information in the surveillance video image can be mined,and the specific human target can be automatically identified,which can save a lot of man power and material resources.At the same time,it can also gain more time to solve the case,and improve the rapid processing ability of the event.There are mainly two research branches:open-world person re-identification and closed-world person reidentification.Open-world person re-identification is in an unknown space environment,and the gallery does not necessarily contain the pedestrians to be retrieved.It is regarded as a sub problem of image retrieval,which is a more challenging and practical application research than closed-world person re-identification,and broadens the application of person re-identification community and research depth.Therefore,in order to solve the above key problems,this paper proposes a person re-identification automatic annotation algorithm based on visual attention mechanism.The main research contents include significant pedestrian detection based on visual attention mechanism,pedestrian automatic annotation algorithm design and construction of semi-supervised person re-identification system.Firstly,the visual saliency region estimation based on Laplacian pyramid fusion is used to quickly estimate the approximate position of pedestrians in the original image or video sequence,and the pedestrian saliency region is obtained,which reduces the pedestrian search range.Then,the pedestrian saliency region is input into the pedestrian detector to quickly obtain the labelling information of pedestrian bounding box and its non line ID category.Secondly,using dynamic sampling unlabeled data pseudo-label algorithm design,according to the prediction confidence,select some reliable pseudo label data from unlabeled pedestrian ID data to construct a pseudo label data subset,with the help of labeled data and index labeled data subset for joint learning,train the automatic pedestrian ID classification model,realize the automatic annotation of person reidentification dataset.Thirdly,the person re-identification system based on semi-supervised learning is constructed by combining the pedestrian detection algorithm and automatic annotation algorithm of visual attention mechanism,which further makes up for the lack of pedestrian labelled samples and realizes the practical application of person reidentification system in real scenes.In order to prove the effectiveness of the person re-identification automatic annotation algorithm based on visual attention mechanism,the performance tests of pedestrian detection,person re-identification and person re-identification system are carried out respectively.In the pedestrian detection stage,the performance test is carried out on INRIA dataset,and the accuracy(mAP)is 88.2%.In the stage of person reidentification,experiments are carried out on Mars,Mrket-1501 and Duke MTMC-Re ID datasets,and the accuracy rates(Rank-1)are 56.3%,49.2% and 59.6% respectively.Finally,the comprehensive performance of the person re-identification system based on semi-supervised learning is tested on the PRW dataset,and the calculation speed of 0.08 frames per second and the accuracy of 47.8%(Rank-1)are obtained.