Person Search Methods Based On Data Augmentation And Multi-Channel Neural Network

With the popularity of cameras in public places,video surveillance has gradually become a very important application in the field of public safety management and playing a particularly important role in social security.However,monitoring systems in public places are usually arranged with hundreds of cameras,which generate a large amount of video and image data,and it is very difficult to manually search person in massive data.Therefore,this thesis takes intelligent person search as the main research content and proposes a person re-identification method based on random linear interpolation and a person search method embedding similarity ranking loss function for the task of searching people with images and texts.In the research work of this thesis,there are three main contributions:(1)For the image-based person search task(i.e.,person re-identification),the data size is small,and causes that the person re-identification network models are easy to over-fitting.A data augmentation method based on random linear interpolation is proposed.The method performs linear interpolation between different samples to generate a large number of new samples and maintains the basic characteristics of the original samples.Data augmentation is useful for exploring the distribution of neighbors of labeled samples and using prior knowledge to supervise learning to more accurately identify and search individuals.(2)In order to obtain the unified representation of deep features between different modal data in text-based personnel search tasks,combined with cyclic neural network and convolutional neural network,a multi-channel neural network framework is proposed to extract different modal data(images and texts)representation in the unified feature space.(3)To further optimize the ranking loss function in the multi-channel neural network framework,a person search method embedding the similarity ranking loss function is proposed.In the network training process,similarity information between image samples is introduced to more accurately constrain the distance between text description and images and improve the performance of text-based person search method.In this thesis,the dataset Market1501 and Duke MTMC-re ID are used in the image-based person search task.On the dataset Market1501,the Rank-1 accuracy of the proposed method reaches 92.71%,and on the Duke MTMC-re ID,the Rank-1 accuracy is reached 82.19%.In the text-based person search task,we use the standard text person search dataset CUHK-PEDES.The Top-1 accuracy of the proposed method is 2.02% higher than the benchmark model's Top-1 accuracy.In addition,the experimental results show that both of proposed methods have good robustness and portability.They can be embedded in the same type of neural network model and have a stable improvement effect on network optimization.