Research On Person Re-identification Algorithms Based On Multi-task Learning

Person re-identification is a key technology in the field of intelligent video surveillance,which can capture and locate targets according to pedestrian characteristics in a large number of pedestrians.At present,person re-identification based on convolutional neural network has been developed rapidly,and many researchers have devoted themselves to the research from the perspectives of data,network structure,loss function and so on.This dissertation draws lessons from the previous researches and makes some improvements and innovations by using the multi-task learning mechanism.The main contributions of this dissertation include the following aspects:1.This dissertation proposes a multi-task learning method based on cross-dataset learning.On the one hand,when the trained model is used for cross-domain testing,the generalization of the model is usually poor and the recognition performance is low.On the other hand,there are few studies on cross-data learning using multi-task learning mechanism.Therefore,this dissertation designs a model of multi-task learning mechanism for cross-data learning,the model is divided into two sub-classification tasks.Finally,experiments show that the model has good recognition ability and strong generalization ability.2.A multi-task learning method based on label uncertainty of human component is proposed in this dissertation.we propose that some components have label uncertainties and the previous training methods for human component network have shortcomings.Therefore,this dissertation innovatively design a network and propose a method of introducing flexible labels to deal with the label uncertainties by using multi-task learning mechanism.Finally,the effectiveness of the improved method is verified by experiments.3.From the perspective of loss function,a multi-task learning method based on global distance scale loss function is proposed.The study of loss function based on distance scale often rely on the constraints between two samples within or between classes,but neglect the integrity.In this dissertation,a loss function based on global distance scale is proposed innovatively.It is more universal by using the statistical properties of distance between classes.It can effectively prevent the interference of outliers or noise and enhance the generalization ability of the model.Finally,by using multi-task learning mechanism and designing two kinds of models,the effectiveness of the models and the ability of loss function are verified.