Attention-Based Pedestrian Attribute Recognition

In recent years,facing at the increasing multimedia data of modern urban monitoring system,including image,audio,video,etc.,how to deal with mass data efficiently has become one of the urgent problems to be solved.In particular,how to accurately analyze the pedestrian image and then quickly identify people and things which are harmful to people's lives and property is the key to improving urban security.Pedestrian attribute recognition task can obtain the attributes of pedestrians according to the pedestrian image.Most existing methods of pedestrian attribute recognition regard it as a multi-classification task of the image and regard each attribute as an independent part,ignoring the correlation between different attributes.In this dissertation the pedestrian recognition task is regarded as a sequence generation task,which makes full use of the relationship among attributes of prediction.Specifically,in order to better learn the mapping relationship between pedestrian images and attributes,this dissertation introduces the attention mechanism to guide the model to pay attention to different parts of pedestrians according to different attributes,the specific work is mainly as follows:The dissertation first proposes a pedestrian attribute recognition algorithm based on image-attribute reciprocally guided attention network(IA~2-Net).According to the feature of pedestrian attributes,pedestrian attributes can be divided into local attribute and global attribute.IA~2-Net firstly uses image features and attribute features to design two kind of guided features:image guided features and attribute guided features,respectively,thus guiding the model to learn the features of pedestrian images according to attributes.Then,in order to preferably make the model give different weights to the two guided features according to the attribute type,a fusion attention mechanism is proposed.Finally,IA~2-Net proposes a novel cross-entropy loss to alleviate the problem of imbalance of pedestrian attribute.Extensive experiments on benchmark PETA dataset and RAP dataset prove the effectiveness and advancement of our approach.In addition,this dissertation also proposes a multi-timestep attention network(MTA-Net)for pedestrian attribute recognition.Firstly,MTA-Net constructs two joint embedding features by using pedestrian images and attributes,and then cascades them with attribute features to obtain fusion features.The fusion feature can guide the model to effectively learn the mapping relationship between pedestrian attributes and images,thus increasing the learning ability of the model.In addition,the existing attention-based pedestrian attribute recognition methods only employ the attribute of current time step to optimize the image feature,leading to ignoring the influence of the attribute of the next time step on the model.MTA-Net proposes a multi-time attention model,which can simultaneously optimize the model with the attribute of current time step and the attribute of next step.Finally,in order to alleviate the impact of the imbalance of pedestrian attributes,we increase the model's attention to the attributes which are difficult to identify,a novel balance loss is proposed.Extensive experiments on benchmark PETA dataset and RAP dataset fully demonstrate the effectiveness and advancement of MTA-Net.