Unsupervised Cross-domain Person Re-identification Based On Multi-scale Features

Cross-domain person re-identification(Re-ID)aims at cross-camera pedestrian retrieval under different domain by computer vision technology,which has a promising future in intelligent analysis of massive surveillance videos in security field,and has been widely concerned by academic and industrial community.There are three issues that should be tackled in real-world Re-ID system,data without label,continuously acquired input data and model's time-efficiency.Existing ReID models focus on solving unlabeled data issue with unsupervised learning,without taking input type of continuously acquired data into consideration,which make model lack of learning ability for new-coming data.Simultaneously,existing models are designed complexly to get high performance,which gives rise to models' processing time and violates time-efficiency criterion.To solve the problem of data without label and poor model's time-efficiency in crossdomain Re-ID,a Re-ID model EMFF based on multi-scale feature fusion is designed from the perspective of improving the effectiveness of feature extracted from pre-trained model.EMFF model builds feature pyramid based on residual network,extracts pedestrian feature of different scales,uses an improved feature fusion method to fuse features,makes fusion feature contains more discriminative information,and improves the model's cross-domain precision.Based on Market-1501 and Duke MTMC-re ID datasets,cross-domain experiments were conducted.On the premise of not using the data from target domain,the EMFF model reached 56.7% and 49.8% on two datasets on CMC Rank-1,and 40.3% and 40.4% on m AP criterion respectively,which were 13.7% and 23.7% higher than Res Net50 benchmark network.It shows that the method can effectively extract the discriminative characteristics of pedestrians and improve the generalization ability of the model.The time-efficiency of the model is tested on the Market-1501 dataset,and the retrieval time of each image is 0.0170 s,indicating that the model's complexity is moderate and can meet the real-time requirement.As for the problem of lack of learning ability of cross-domain Re-ID model,ILReID is designed based on unsupervised clustering and incremental learning.Based on unsupervised clustering learning,pseudo labels are added to unlabeled data,and pseudo labels are used to train the pre-trained model.In order to simulate the data input mode in the real scenarios,the datasets were divided into several non-overlapping sequences of datasets,and the model was trained in sequence using different datasets based on unsupervised incremental learning.The model is updated online based on the method of parameter adjustment.By conducting cross-domain experiments on Market-1501 and Duke MTMC-re ID datasets,the model reached 82.5%? 75.1% respectively on CMC Rank-1,and 59.1%? 57.2% respectively on m AP,which increased by 0.8%? 3.8% respectively compared with the existing optimal methods based on unsupervised domain adaptation.It shows that the method can have a better domain adaptation on the target domain through incremental learning,and the model has the ability to learn new data,and can effectively deal with the scene of real-time acquisition of data in the monitoring system,which is more suitable for practical deployment.