Design And Implementation Of Active Learning And Condition Matching Person Re-Identification System For Winter Olympics Town

In order to enhance the service capability of the intelligent operation platform for the Winter Olympics smart town,which encompasses various scenarios and cross-industry operations,it is proposed to digitally analyze and model the town’s resources.In order to build a digital model of bus passengers,this thesis analyses pedestrian pictures based on person reidentification algorithms and achieves the identification of pedestrians riding on the bus.Further,this thesis designs a condition matching person re-identification approach,which can correlate pedestrian identity and pedestrian boarding and alighting actions to achieve the tracing of pedestrian ride trajectories.Considering the widespread problem of data distribution bias in the deployment of deep models for person reidentification,this thesis investigates active learning-based person reidentification on open datasets to correct the distribution bias between the deployment domain data and the training domain data with a small sample size.Furthermore,to effectively evaluate the effectiveness of active learning in person re-identification and to reduce the manual annotation operations during active learning evaluation,this thesis designs an active learning simulation testing framework for person re-identification.In addition,this thesis proposes a microservices-based design and architecture of a distributed system for person re-identification,optimises the problem of difficult authentication of interface permissions in the microservices architecture,and implements the aforementioned framework.The person re-identification system is based on an object-oriented design,with independent modules designed for the steps of data sampling,data annotation and model evaluation in active learning person re-identification and satisfying the LSP,which can support active learning person reidentification and application research.1.A condition matching person re-identification framework that can combine the results of person re-identification and pedestrian boarding and alighting action detection is designed.This framework is implemented based on distributed microservices and can be used to retrieve pedestrians with specific boarding behaviors..2.An active learning framework for person re-identification is designed.Considering the data distribution offset problem of open-world person re-identification,this thesis investigates the iterative mechanism of a distributed active learning-based person re-identification model.Compared with random sampling,the performance of person reidentification models under active learning conditions is improved faster and the amount of manual annotation is saved.3.An active learning person re-identification simulation testing framework is designed.Based on the active learning simulation test method that masks part of the true value of the dataset,the thesis designs an iterative person re-identification model that can simulate a humanin-the-loop without repeated manual annotation,which can be used to evaluate part of the performance of the active learning algorithm.4.A microservices-based active learning and conditional matching person re-identification system is developed.To protect the interface security,this thesis implements an RBAC hot-reload responsive API gateway,which provides different OpenAPI permissions to users with different identities and optimises the performance of the gateway through responsive programming and distributed caching.During system deployment,this thesis uses a devops approach to simplify microservice operations,which helps agile development and rapid iteration.Moreover,the business and algorithms are implemented and deployed in multiple microservices separately.Compared with the monolithic architecture,the proposed solution has better scalability and extensibility.