Research On Person Reidentification Task Based On Monitoring Scenarios

The purpose of Person Re-identification is to find all picture of a pedestrian from the database containing this pedestrian given a picture of it,which has always been an urgent task in the field of computer vision.Although the person re-identification technology has achieved good results on a single public dataset,the scale of the public dataset is not enough for the trained neural network to do the person re-identification in the case of large-scale pedestrians under the real monitoring scene.At the same time,in the real scenarios,we often need to face the problem of large-scale data update.Finally,person re-identification technology cannot solve the whole security tracking task independently.Therefore,this thesis contains tasks such as unsupervised domain adaptation person re-identification,lifelong learning person re-identification,and multi-target multi-camera tracking based on deep learning methods.The details are as follows:For the problem of domain adaptation person re-identification,this thesis proposes a method based on deep mutual learning,in which the influence of pseudo-label noise on model accuracy is reduced by using the construction of pseudo-labels in the camera.It is proposed to make full use of the camera information and timestamp information in the dataset to further improve the accuracy of the model.Secondly,this thesis proposes a life-long learning person re-identification framework,by using KL divergence and classification loss to optimize the new task without forgetting the old task,and proposes an improved loss function field to adapt to the negative sample loss function to increase the distance between the new and old task features in the feature space,thereby it improving the overall accuracy of the lifelong learning person reidentification task.In order to face the real scene,based on a real monitoring scenario,this thesis constructs a dataset,so as to adapt the situation of real scenarios application.Finally,this thesis proposes a multi-target multi-camera tracking framework that combines person tracking technology and person re-identification technology to compare the average feature of pedestrian trajectories from multi-camera,supplemented by a camera connectivity graph generated by counting the connectivity probability between cameras in real data,based on the fact that the same pedestrian tag cannot appear under different cameras at the same time,the range of trajectory tracking is further constrained.The range of trajectory tracking is constrained,thereby improving the overall accuracy of the multi-target cross-camera framework.