Pedestrian Attribute Recognition In Security Scene Based On Deep Learning

China's modern security construction has been developing for decades.From the initiative of 'digital city' to the advocacy of 'smart cities' based on the IoT and artificial intelligence,security construction has gone rapidly on the road of modernization and intelligence.Due to the construction of a huge security network,the amount of data that needs to be processed is large,the traditional method of storing data and then analyzing it is not real-time and inefficient.Pedestrian attribute recognition technology based on deep learning can effectively combine modern monitoring networks,real-time analysis and quickly locate suspicious characters,which helps to cope with problems in the field of security,and accelerate the construction of smart cities.Pedestrian attributes have a broad meaning.Pedestrian's gender,age,clothing style,and pose are all pedestrian attributes.Pedestrian attribute recognition belongs to the category of computer vision technology which achieves attribute recognition by converting low-level semantic information to medium-level semantic information including feature extractor to extract features,classifier to do the classification and others.In traditional machine learning algorithms,features are designed by researchers,but in the task of pedestrian attribute recognition,due to the darker brightness and occlusion,it is difficult to design features manually and the generalization performance is bad.Deep learning technology finds the necessary features by the model,and can mine the connections between pedestrian attributes.Therefore,when dealing with the actual pedestrian attribute recognition task,the deep learning method is more robust.This paper first introduces the research status of pedestrian attribute recognition technology,then explains the main technologies used in it,and finally uses a residual neural network based on multi-attribute fusion training idea for pedestrian attribute recognition experiments.The 50-layer residual neural network innovatively adopts a bottleneck structure which makes it easier for the shallow layer of feature information to be transmitted to the deeper level.Therefore,the network based on bottleneck structure has more layers and can process more complex information.Selecting the residual neural network as the backbone network allows the trained model to have a better recognition performance when faced with complex environments and angle changes in security scenarios.In the first part,the paper defines the scaling parameters for the weighted sigmoid cross-entropy loss function to study the effect of scaling parameters adjustment on pedestrian attributes with different positive rates.The experiment using a 50-layer residual neural network and the processed RAP2.0 dataset.The dropout setting is canceled to reduce randomness.Precision,recall rate and F1 score are used to judge the experimental results.It is found that within a certain range,increasing the scaling parameter will reduce model's recognition performance of the attribute with a higher positive rate on the recall rate,precision and F1 score,and improve the performance of the attribute with a lower positive rate on these three indicators,while reducing the scaling parameters will have the opposite effect.It provides an experimental basis for us to improve the balance of the model's recognition performance at the expense of model's recognition performance of the attribute with a high positive rate in pedestrian attribute recognition projects.The second part of the experiment establishes the process of expanding the model's recognizable attributes in the actual attribute recognition project.By using the correlation between the attributes and expanding the dataset for the new attribute in a short period of time,we improve the model recognition accuracy.This part uses 50-layer residual neural network for experiment.In the experiment,the shoe style attribute dataset was expanded through this process.Firstly,among the six attributes of age,top style,top color,trousers style,trousers color,shoe color,shoe style,shoe color attribute had the strongest correlation with shoe color.By training two strongly related attributes jointly,the model pre-annotated accuracy is improved,the time required for manual cleaning is reduced.Within a week,after 5 iterations of the process,2600 shoe images and corresponding tags were expanded.After training with new dataset,the accuracy of the shoe styles was improved by 2.7%.It is proved that the process I proposed utilizes the strong correlation between attributes and improves the efficiency when expand dataset of the new attribute.The system used in the experiment is Ubuntu,the model building framework is PyTorch.The experiment uses the RAP2.0 dataset,which data is composed of actual security images and labels and has many marking attributes,which meet the experimental requirements.