Activity Feature-Based Human Identification

Human identity recognition has shown great practical value in realworld applications,and it has very important research significance.This thesis studies the problem of activity-based human identification from two perspectives of metric learning and person re-identification.Metric learning is also called similarity learning.In the field of image recognition,it is usually used to compare whether two pictures are the same object.In the field of identity recognition,it is used to judge whether two human pictures are the same person.Metric learning has been studied in the field of human gait recognition for more than ten years,which is due to the good characteristics of human walking posture,such as pedestrian features can be obtained without actual interaction with pedestrians.However,people's walking images are easily affected by a variety of factors such as resolution,view and clothing,which may lead to the poor performance of distance measurements.In practical application,human may perform other activities.To this aim,this thesis studies identity recognition from human activities by employing metric learning methods based on Mahalanobis distance.Firstly,background subtraction is used to extract human contour,and then principal component analysis is used to reduce dimension.When calculating the distance between features,this thesis exploits CSML,KISSME,GMML and SILD for comparison.Person re-identification aims to solve the problem of pedestrian retrieval in multi camera environment.The technical difficulty of deep convolution neural network is how to use the network model to fully mine and extract effective human features,so that the extracted pedestrian features can be discriminative.In this thesis,a new network structure is designed and a self-adjusting feature learning(SAFL)method is proposed to alleviate the problem that the network model has limited ability to learn human feature representation.The SAFL can make use of the relationship between global features and local features.By introducing the dynamic self-adjusting method,it can realize the mutual supervision between the features of network learning to deeply mine the potential features,so that the network can obtain a more discriminative pedestrian feature representation.The proposed methods are compared with baseline methods on datasets of human identity recognition and pedestrian re-recognition,and the experimental results show the effectiveness of the proposed methods.