Research On Pedestrian Attribute Recognition Technology Based On Deep Convolutional Neural Network

In the big data era,the concept and construction of smart cities are flourishing.As a security guarantee for a harmonious society,surveillance cameras can be seen everywhere.Pedestrian attribute recognition under surveillance video realizes the structure of pedestrian information through statistical integration of information,which is of vital importance and significance in the fields of intelligent security,criminal investigation,and accurate advertising.However,due to the complexity of surveillance video scenes,and the correlation between pedestrian attribute categories and the association between attribute categories and spatial locations is difficult to mine,the task of pedestrian attribute recognition is still challenging.Pedestrian attribute recognition method based on semantic segmentation is proposed in this paper.Designed a multi-branch network structure based on visual attention,for the task and data characteristics of pedestrian attribute recognition.Taking full advantage of the advantages of semantic segmentation,the visual attention of attribute semantic information and the spatial attention generated by human parsing are effectively fused.Specifically,for the problem of low general target detection efficiency,a real-time pedestrian detection algorithm is designed,which achieves a good balance between speed and accuracy.In terms of pedestrian tracking,a more stable online and realtime tracking method of a deep association metric is used.In order to improve the image quality in the pedestrian library,the image of the same pedestrian target is chosen using a general image quality assessment method.In terms of pedestrian attribute recognition,in view of the uneven distribution of data in pedestrian attribute recognition and the association between attributes that are difficult to mine,a multitask loss function with weights is proposed,and a multi-branch based on visual attention mechanism residual network is designed.In order to make better use of the relationship between pedestrian attribute categories and spatial location distribution,the semantic segmentation auxiliary information is introduced.The human parsing network combines the semantic feature expression of segmentation maps and the boundary perception features of edge maps.A self-correcting iterative training strategy is used to train the model.The fusion method of human parsing segmentation maps and attribute recognition feature maps is proposed,which effectively applies the spatial information brought by human parsing to the attribute recognition network This method achieves a strong association between attributes and between attributes and spaces,which has achieved a high accuracy rate of 83.4% in the benchmark data set WIDER Attribute.