Clustering-Based Unsupervised Bus Passenger Re-Identification

An intelligent public transport system is of great significance to bus scheduling and line network optimization.Timely and accurate analysis of passengers’ movements can make travel safer and more convenient and effectively improve the quality of bus services.Presently,pedestrian re-identification technology has achieved extensive research results in the academic field,while the model accuracy will significantly reduce for scenarios such as bus passengers with severe occlusions.Moreover,supervised pedestrian re-identification relies on manual labeling,which is timeconsuming and labor-intensive,and using unsupervised methods can save the cost of manual labeling,so this paper conducts a study on unsupervised re-identification of bus passengers based on clustering.The main work content is as follows.(1)Unsupervised bus passenger re-identification algorithm based on clustering.This chapter uses the unsupervised domain adaptation method for training.The source domain is public dataset with identity labels,and the target domain is unlabeled bus passenger head dataset.Two operations are alternated.First,the network pre-trained on the source domain extracts features for the data on the target domain and uses feature clustering to generate pseudo-labels.Second,use pseudo-labels to supervise network learning on the target domain.By experimenting with different clustering methods,select a better clustering method and the number of clustered pseudo-classes to provide data support for subsequent work.(2)Bus passenger re-recognition algorithm based on collaborative average training.The pseudo-labeling noise generated by the clustering-based unsupervised rerecognition algorithm in the clustering stage and the limitations of the clustering method hinders the ability of the model further improve the feature representation on the target domain.In order to solve the above problems,this paper proposes a collaborative average training model for bus passenger head re-recognition task and uses a soft maximum triplet loss to adapt to soft pseudo-labels.By experiments,m AP and rank-1 on Market-to-Head increased by 7.6% and 7.8%,respectively,and 6.1% and7.4% on Duke-to-Head compared to the unsupervised clustering-based algorithm.