Unsupervised Vehicle Re-Identification by Spatio-Temporal Contrastive Learning

Vehicle re-identification (Vehicle-ReID) is one of the key technologies for intelligent monitoring of traffic network status. It can identify the same vehicle in different fields of view by relying on cameras with non-overlapping fields of view, so as to infer the running path of the vehicle in the road network, and pave smooth way for various applications, such as multi-view vehicle tracking, abnormal diagnosis and road management and control. Most of the traditional Vehicle-ReID methods rely on large-scale human-annotated ID labels, which undoubtedly affects the cross-scenario generalization ability in practical use. To this end, this paper proposes an unsupervised vehicle re-identification method based on spatio-temporal contrastive learning. It mines the easily collected camera locations and the timestamp of vehicles and combines them with the deep features of the vehicle image. Then, a multi-view fusion feature dictionary is formed by the feature clustering approach, where the multi-view feature dictionary is used to compute the estimated loss with the feature vector of the query vehicle image, and updated incrementally to optimize the feature dictionary. Compared with the existing state-of-the-art unsupervised Vehicle-ReID methods, the proposed method outperforms them through exhaustive experiments. Furthermore, this paper constructs a new large-scale vehicle re-identification data platform HvRi, which contains 12,456 images of 784 vehicles from 99 cameras over 24 hours. The scale of cameras and the number of vehicles are much larger than existing public datasets. Among them, the model achieved the rank-l rate of 88.6% on the public VeRi-776 dataset (20 cameras, 776 vehicles), which has a high engineering application value.