Pedestrian Attribute Recognition Based On Attribute Localization Positioning And Relation Mining

Pedestrian Attribute Recognition（PAR）can accurately identify the identity and appearance information of pedestrians.PAR is widely used in the field of intelligent video surveillance,which is very important for security inspection,recognition and business intelligence based on video.Therefore,PAR has a very high development prospect.PAR can extract pedestrian appearance information,which can extract a variety of different information according to given images and data.PAR can help people better understand the personality and appearance of pedestrians and provide more accurate information.At present,the technology is at the forefront of computer vision,and PAR technology is also facing problems such as single pedestrian characteristics,pedestrian posture change and occlusion.This thesis uses computer vision and deep learning to study the above problems.Specific work is as follows:（1）Research on Pedestrian Attribute Recognition based on multi-level feature fusion.A variety of attributes of pedestrian images need the output features of different network levels,and a single feature can’t predict each attribute well.This thesis proposes a Multi-level Feature Fusion（MFF）network.BN-Inception network is used to extract shallow features,middle features and deep features,and then feature fusion module is used to fuse the three levels of features,so as to improve the prediction accuracy.In order to solve the aliasing effect in feature fusion module,this thesis proposes a new residual connection module,which adopts attention mechanism and can effectively enhance features.The MFF network achieves 87.20%and 89.40%on the PETA dataset m A and Recall respectively,which is 1.28%and 3.17%higher than the advanced model SSC_hard.The Recall and F1 of PA100K dataset reach89.40%and 86.92%respectively,which is 1.85%and 0.37%higher than the advanced model SSC_hard.（2）Research on Pedestrian Attribute Recognition based on attribute location.Due to the random change of pedestrian posture,the relative position of attributes is uncertain.To solve this problem,this thesis proposes an Attribute Localization Positioning（ALP）module.It can accurately identify the attribute region and extract the feature representation of the attribute-related region by using attribute location.In this thesis,the ALP module is integrated with MFF network,and an Attribute Location Based on Multi-level Feature Fusion（ALMFF）network is proposed.On the PA100K dataset,ALMFF network achieved 81.89%,78.74%,89.97%and 87.01%on the indexes of m A,Accuracy,Recall and F1,which were improved by 0.87%,0.32%,2.32%and 0.96%respectively compared with the advanced model SSC_hard.（3）Research on Pedestrian Attribute Recognition based on attribute relation Mining.In order to solve the occlusion problem of attributes,this thesis proposes an Attribute Relationship Mining（ARM）network.The ARM network is divided into Feature Enhancement（FEN）module and Attribute Correlation Mining（ACM）module.the FEN module mainly deals with pictures,and the attribute relationship module deals with label characters.The core of ACM module is the construction of correlation matrix,which is a static matrix based on the whole dataset.Secondly,a classifier is generated through GCN processing to classify the image features.In this thesis,the modules in MFF network and ALP module are added to ARM network,and the final complete network,Attribute Location and Relationship Mining（ALRM）network,is proposed.The Accuracy,Precision and F1 of ALRM network on RAPv2dataset reached 71.24%,83.10%and 82.48%,respectively,which were improved by3.21%,4.35%and 2.79%compared with the advanced model IAA-Caps.