Research On Semi-Supervised Person Re-Identification Based On Single Camera Annotation

With the continuous development of camera monitoring technology,person re-identification research has received great attention and a lot of exploration.Person re-identification utilizes computer vision technology to search for target pedestrians across different cameras.Given an image of a target pedestrian,the goal is to retrieve the images of the same pedestrian captured by cameras in different regions.Due to the significant value of person re-identification in social security,its application in large-scale surveillance systems is a trend for future development.Most existing person re-identification methods focus on small-scale surveillance systems,where each pedestrian is captured in different camera views in adjacent scenes.When person re-identification scales to larger systems,it requires pedestrian matching not only across adjacent scenes,but also across distant scenes.For example,a criminal suspect is on the run in a different city.However,the lack of cross-distance identity labels in large-scale surveillance systems covering larger areas poses a huge challenge for person re-identification research.In this thesis,we study person re-identification systems based on single-camera labeling,and discuss adaptive joint learning person re-identification and camera feature learning person re-identification,respectively.The main research contributions are as follows:(1)In this method,each identity is annotated in a single camera.Since there are no identity annotations across cameras,the time required for data collection is significantly reduced,and rapid deployment in new environments can be achieved.To address the problem of single-camera person re-identification,a joint contrastive learning framework is proposed,dividing the training data into three parts:single-camera labeled data,pseudo-labeled data,and unlabeled instances.In this framework,the network is iteratively trained,and the memory is dynamically updated to store three types of data.Cluster algorithms are used to assign pseudo-labels to unlabeled images.Unlike traditional contrastive learning,this framework jointly distinguishes single-camera labeled data,pseudo-labeled data,and unlabeled instances.Through comparison experiments on three large datasets of person re-identification,it is verified that the joint comparison learning framework proposed in this thesis can effectively improve the performance of person re-identification models.(2)This method uses a semi-supervised setting with single-camera annotations,where the training data includes single-camera labeled data and unlabeled data.In person re-identification across distant scenes,a feature prediction algorithm is proposed to supplement the missing cross-camera pairing data.Camera information is then introduced,and a triple loss for learning camera information is proposed,which not only learns similar features between cross-camera anchors and positive samples but also learns distinctive features between intra-camera anchors and negative samples.To effectively utilize a large amount of unlabeled data,each individual in the unlabeled image is considered an independent entity,and diversity and similarity losses are used to train the unlabeled data.This thesis shows through extensive experiments that the proposed method performs better compared to other in-camera supervised person re-identification algorithms.