Person Re-identification For Robust Feature Exrtactionp In Complex Scenes

Person re-identification is one of the key points of intelligent surveillance research.It uses computer vision related technologies to determine whether the same person exists in non-overlapping areas between different cameras,and can then achieve rapid retrieval of suspicious people and predict related abnormal events.However,in practical application scenarios,due to the problems of lighting changes,background clutter,various pedestrian poses,and occlusions,there are many challenges in person re-identification research,which have caused widespread concern in the academic and industrial fields.How to extract more robust and discriminative pedestrian features to improve person re-identification has become one of the core research topics of this subject.It is difficult to eliminate the impact of the above problems by using the hand-crafted features.In recent years,it has been found that the use of convolutional neural networks can extract more discriminative image features and use them for person re-identification has also improved significantly.This article is inspired by human visual observation rules and learning knowledge methods,and explores person re-identification methods based on attention-guided interaction and curriculum learning strategies.First,the network model combines the attention mechanism to generate an attention map in a supervised manner.The attention map is applied to the global and local branches,taking into account global and local information,and using local features to strengthen important areas in the global.Use global information to provide contextual information for local information.In addition,define the complexity using density distribution,to re-divide the dataset,use curriculum learning strategies to train models from easy to difficult,and gradually improve network performance with progressive multi-stage training.Through experimental verification on three benchmark datasets of person re-identification,the proposed method can achieve the extraction of robust features of pedestrians in complex scenes.Compared with state-of-the-art,the performance is improved to a certain extent.Especially in the most challenging large dataset MSMT17,our method outperforms the existing methods,Rank-1 and m AP increase by 9.1% and14.6%,respectively.Finally,the validity of our method in practival application scenarios is achieved through a robot platform.