Research On Feature Representation And Metric Learning In Person Re-identification

With the rapid increase in the use of surveillance cameras and monitoring video data,manually process video is time-consuming and grueling.The demand for automatic analyzing and processing video data by computers is required.The person re-identification problem refers to recognize person automatically by computers from surveillance video data with non-overlapping camera coverage.It is a practical application of monitoring video data mining.The data needed to be processed varies greatly due to many factors such as multi-camera,the differences of shooting environment,the changes of pedestrian's appearance and posture,and so on,so person re-identification is considered as a fine object recognition problem.At present,there exists a large gap between the research on person re-identification and the actual landing application.This article summarizes several research hotspots of person re-identification in the current stage,and launches research in pedestrian feature extraction and distance measurement methods.The main research work is as follows: 1)Research on deep network model based on QRNN for local pedestrian feature extractionPedestrian recognition is a fine object recognition problem,and the extraction of local detail features is more important.This paper supposes that the image of pedestrian along vertical direction is viewed as a timing signal,and bi-directional QRNN,an improved QRNN,is used to extract the local details of pedestrian images.The network structure is trained by multi-classification loss function.Experiments show that the person re-identification algorithm based on QRNN can obtain good experimental results.2)Research on deep network model based on joint learning of multiple loss functions for pedestrian feature extractionThe choice of loss function has an important impact on the performance of person re-identification algorithms based on deep network architecture.This paper introduces a multiple deep network model fusion strategy DML algorithm to solve the problem of inconsistent label probability of the two-branch softmax loss function in the JLML model,which helps to improve the feature extraction capability of the deep network model.At the same time,a comparative analysis of various loss functions commonly used in person re-identification algorithm is proposed.Aiming at solving the difficulty of training the triplet loss function encountered during the experiment,a multi-loss function stepwise training strategy is proposed.3)Person distance measurement algorithm based on Equidistance Measurement LearningMany metric learning methods in person re-recognition problem define the distance difference between similar sample pairs and heterogeneous sample pairs by “interval” hyper-parameters.This paper introduces the idea of equidistant metric learning into MLAPG algorithm to avoid cross distribution of positive sample pairs and negative sample pairs in the transform space when algorithm convergences,which improves the performance of person re-identification algorithm significantly.