Pedestrian Attribute Recognition Based On LSTM

With the wide application of monitoring system in the city,more and more attention has been paid to the research of intelligent monitoring system.Especially,the detection and recognition of pedestrian under surveillance video has become the focus and challenge in the field of intelligent surveillance,and the automatic identification of pedestrian attributes has become an increasingly important issue.One common problem is that the monitor wants to search for the specified pedestrian according to the characteristics of the target.Before the emergence of an intelligent surveillance system,the monitor can only search the target with eyes in the monitor screen.But with the increasing number of monitoring probes,massive surveillance video will undoubtedly bring great challenges to monitors.So automatically identifying the fine attributes of the person in the monitor picture has important practical significance.However,due to the diversity and complexity of the human attributes in the surveillance screen,this task is faced with many challenges.The traditional machine learning methods treat this problem as a pedestrian multilabel classification problem,and use support vector machines or convolutional neural network to predict the pedestrian attributes.Different from these existing methods,this thesis borrows the idea of image caption and transforms the classification of pedestrian attributes into the "generation" of pedestrian attributes,and presents the attribute recognition of pedestrian images using convolutional neural network and recurrent neural network.The method first obtains the attribute contextual information by integrating the pedestrian attribute labels,and then uses the convolutional neural network to extract the image visual features.The contextual information is analyzed by Long Short-Term Memory to get the relationship among attributes.Further,this thesis also proposes a new recognition model based on guiding Long Short-Term Memory networks,which adds the guiding information of attribute description.Experiments on PETA datasets show the effectiveness of the proposed method and outperforms the state-of-the-art methods on the benchmark PETA dataset.