Research And Implementation Of Pedestrian Attribute Recognition Algorithms

Pedestrian attribute recognition is a significant task in computer vision research and plays an important role in video surveillance.In recent years,it has been applied to person re-identification,face verification and person retrieval,etc.Pedestrian attributes are the description of pedestrian semantic features,including clothing attributes and biological features(e.g.age and gender).Although there have been a lot of researches on pedestrian attribute recognition,it's still a challenging and pioneering problem in the computer vision research due to the influence of occlusion,blur and low resolution.In this paper,pedestrian attribute recognition is modeled as a multi-label classification task,and there are correlations between attribute labels,for example,people who wear dress are more likely to be female,and most male don't have long hair.Therefore,a model based on dictionary learning is proposed in the paper to characterize the correlations of attributes effectively.Furthermore,this paper models the correlation between attribute labels as a nonlinear structured relationship,and proposes an end-to-end learning model based on fusion convolutional neural network and graph convolutional network,so as to improve the recognition performance.The main works of this paper include the following two contents:(1)Pedestrian attribute recognition model based on dictionary learning.In order to model multiple attributes jointly,the attributes are modeled as a subspace and a dictionary is introduced to represent the subspace,where each atom of the dictionary stands for an attribute.Specifically,the dictionary is used to project the features to a sparse vector,and the i-th element of the vector corresponds to the predicted confidence of i-th attribute.At the same time,in order to learn the features that are more suitable for the attribute prediction,the dictionary is modeled as a network layer learning with convolutional neural network jointly,which is called attribute prediction layer.Finally,according to the learning methods of dictionary and convolutional neural network,a joint optimization algorithm based on dictionary and convolutional neural network is proposed.(2)Pedestrian attribute recognition model based on fusion graph convolutional network.The dependency between attribute labels can be modeled not only as a linear relation,but also as a nonlinear relation.Therefore,in order to capture and explore these nonlinear relations,an end-to-end attribute recognition model fusing graph convolution network is proposed in this paper.In this model,each node represents an attribute whose features are learned through the convolutional neural network.Furthermore,according to the probability of each pair of attributes appearing simultaneously in the training samples of the dataset,the relation matrix of attributes is built to guide the information propagation among the nodes in graph convolution network.At the same time,in order to increase the generalization of the model and reduce the complexity of the network,the threshold is adopted to optimize the relational matrix.To model the correlations between attributes,two attribute recognition algorithms are proposed in this paper.Experimental results show that the algorithms improve the accuracy of attribute recognition on the mainstream pedestrian attribute datasets PETA and PA-100 k,and achieve state-of-the-art performance.