Design And Implementation Of Cross Modality Person Re-identification Algorithm

With the continuous development of today's society,camera equipment is constantly recording and monitoring the steady development of society.In the smart cities that are currently receiving much attention,smart surveillance also forms an important part of security in smart cities.With the gradual promotion of face recognition technology,person reidentification technology has been studied by a lot of scholars and institutions because of its wide application scenarios and research significance.Person re-identification has great potential in areas such as smart surveillance and urban security.With the iteration and optimization of deep learning algorithms,and the emergence of many large public datasets related to person re-identification,person re-identification under the lens of visible light has reached a relatively high accuracy.However,compared to daytime,night surveillance can only rely on infrared cameras to obtain clearer images.In fact,infrared surveillance plays a big role in urban security and indoor surveillance.However,there are relatively few related researches on infrared-visible light cross-modal person re-identification for night surveillance,and most of them have difficulty in solving the cross-modal difference between infrared and visible light images,and there is still much room for improvement in this field.Aiming at the problem of poor recognition accuracy of infrared-visible cross-modality person reidentification,this thesis mainly makes the following researches:First,the existing work mainly focuses on using two models to extract the features of infrared and visible person images,respectively,and then perform classification and metric learning through a fully connected layer that shares weights and the same loss regression layer.However,through experiments and analysis,we found that using the two models to extract the features of the two modalities separately,the fitting effect on the features is not as good as using the models with shared weights for feature extraction.Compared with cross-modal recognition such as text-image or text-audio,the similarity between infrared and visible light in two dimensions is relatively high,so the alignment effect after converting to one-dimensional feature values is not ideal.This thesis proposes a cross-modal person reidentification model based on residual networks and local features.This model is a simple and effective end-to-end cross-modality person reidentification model.By focusing more on the low-level features of the model,the accuracy is improved under the premise of increasing the amount of model parameters.Secondly,in order to further reduce the feature difference between the two modes of infrared and visible light,this thesis uses the Euclidean distance based mean square error loss constraint feature,and simultaneously proposes a multi-layer loss constraint function to enhance the model for common features when extracting low-level features Extraction.At the same time,this thesis uses triangle learning rate to further promote model convergence.Finally,although the current public dataset for cross-modal person reidentification has expanded from thousands of pictures to tens of thousands of pictures,compared to the dataset of tens of millions of pictures for face recognition,the cross-modality person re-identification dataset is still too small,so this article further amplifies the data set by randomly erasing some areas on the picture to enhance the robustness of the model.Experiments show that our model achieves the first hit rate Rank1 =70.20% and the mean Average Precision m AP = 54.21% in the full-scale multiple search mode of the public dataset SYSU-MM01,which is a crossmodal person re-identification.The first hit rate is Rank1 = 86.17%,and the mean Average Precision is m AP = 81.81%,which are significantly higher than the data of existing methods.