Person Re-identification Technology Based On Background Elimination And Attribute Recognition

Person re-identification(person Re-ID)technology has attracted extensive attention from academia and industry recently.While showing great potential,it also faces many challenges,including low resolution,object occlusion and various person posture.At present,person Re-ID mainly focuses on two aspects,person feature representation and distance measurement.The former mainly aims to extract the most representative person feature and the latter designs an appropriate person distance measurement function based on the acquired person feature.This thesis mainly studies how to extract effective person feature.The main work are summarized as follows:(1)Training Res Net50 network with the whole person images and the segmented person foreground images,the effect of person background on recognition performance under the same-dataset and cross-dataset was studied.Experimental results showed that using the whole person images has better performance under the same-dataset,and using the person foreground images has better performance under the cross-dataset.It can be concluded that the network may learn the feature of background while learning the feature of person.Thus,it is necessary to design an effective background elimination scheme adapting to the actual scene.(2)In order to solve the problem of the loss of advanced features in basic Res Net50,a multi-pool fusion(Multi-Pool Fusion,MPF)network structure was developed.Multiple feature vectors were obtained by pooling the global and local feature maps with different pooling types.Thus multiple independent mapping relationships between feature maps and feature vectors were established,which can balance global and local features and implement multiple fine-grained representations.Experimental results showed that MPF has greatly improved person Re-ID performance.(3)To eliminate the influence of background information on person Re-ID,pixel-level background elimination,feature-level unsupervised background elimination and feature-level supervised background elimination were proposed.Pixel-level background elimination method directly removes background information at the input end.In the feature-level unsupervised background elimination method,segmentation masks are used to just retain the feature data of pedestrian foreground images.The feature-level supervised background elimination method combines the feature activation loss function with the person classification loss function to supervise the network model to extract useful person foreground features.Experimental results showed that the feature-level supervised background elimination method has better recognition performance in the same-dataset and cross-dataset.(4)Existing person Re-ID model combined with attributes information ignores that attribute information is related to image region.To solve this problem,an activation guided attribute classification(Activation Guide Attribute Classification,AGAC)model was proposed.The model includes a branch guided person and attribute classification module(BGPACM)and a mutex local activation module(MLAM).BGPACM employs the relevant branch to classify each person attribute,which improves the accuracy of attribute features.MLAM enables each branch to generate mutex activation areas,which can increase the overall network feature representation capability.Experimental results showed that AGAC model has better recognition performance than existing attribute information combined person Re-ID methods.