Person Re-identification Based On Deep Learning And Its Application In Airport

Person re-identification refers to the recognition and retrieval of target pedestrians based on person video data captured by multiple cameras in cross-camera or cross-scenario situations.In recent years,with the rapid development of deep learning and its continuous application in the field of computer vision,the performance of person re-identification has been greatly improved.However,for practical applications,person re-identification still faces great technology challenges: 1)It is difficult to generalize and extend the person re-identification model,which has become a big obstacle that restricts the practical application of person re-identification;2)The insufficient size of the dataset also restricts the further development of person re-identification.Based on deep learning technology,this paper studies and proposes methods to solve the problem of person re-identification from different levels,including:(1)In response to the image sequence feature clutter problem faced by one-shot video person re-identification,a person re-identification model based on a deep discriminative network is proposed in conjunction with a multi-objective function.In this chapter,the following problems are mainly solved: 1)considering the characteristics of one-shot video dataset,each person has only one image sequence data,but it also contains rich feature information,such as time sequence information,different perspective information of the same person,and pose information,etc.,a hybrid loss function algorithm for sequence-bound images is proposed to purify the feature information of pedestrians in sequence images,so as to obtain better robustness of the model;2)in the model training phase,a novel semi-supervised learning sampling strategy for step-by-step learning is adopted,and unlabeled data is gradually used to improve model performance within acceptable error limits;3)in order to make the model extract more discriminating pedestrian features,a random sampling measure with dynamic adjustment interval is proposed.The experimental results show that the proposed model can well solve the problem of feature cluttering of sequence images in video-based pedestrian re-identification tasks,and also solve the problem of effective use of untagged data for pedestrian re-identification based on semi-supervised learning.(2)In response to problems such as the existing datasets of person re-identification are too small for the model to be adequately trained,too little single-sample data makes it difficult to train a robust person re-identification model,and the challenges of effectively using unlabeled samples,the study proposes a person re-identification network model based on data augmentation.In this model,a new progressive sampling algorithm based on semi-supervised learning that combines single sample data and enhanced data is proposed.When the model is initialized,the original data is processed by a novel random data augmentation algorithm.Data augmentation to make the model more fully trained,and then continue to classify the unlabeled data through the classification model,and select the most reliable data and its enhanced data to bring into the new round of model training.Enhancing the data can not only reduce the probability of overfitting the model,but also expand the single sample data to achieve the purpose of fully training the model.In each iteration,as the enhancement data promotes the training of the model,the accuracy of the model in introducing unlabeled data also gradually improves.The final experimental results also show that in each iteration of the model,the accuracy of the pseudo-label data has been significantly improved.(3)Combining person re-identification and target detection algorithms,this paper explores the application of person re-identification on airport specific datasets.The surveillance video with a more cluttered background outside the airport and the surveillance video of the same person taken by two cameras under the same corridor in the airport were selected.The video contains interference factors such as background changes,person attitude changes,and perspective changes.The proposed person re-identification algorithm combined with target detection achieved good results.