View-Decision Based Compound Match Learning for Vehicle Re-identification in UAV Surveillance
Song Ye, Chunsheng Liu, Wang Zhang, Zhaoying Nie, Luchang Chen · 2020
Focus on the vehicle re-identification task in the unmanned aerial vehicle surveillance, the View-Decision Based Compound Match Learning (VD-CML) method is proposed, which embeds image viewpoint information in network structure and metric learning to enhance the ability to re-identify vehicles in different viewpoints. First, the viewpoint judgment model is constructed to obtain the viewpoint information, and the viewpoint decision model generates the composite sample pair embedding viewpoint information. Then, the multi-branch separable Siamese network and the multivariate compound contrastive loss function are designed to learn the feature of composite sample pairs in three viewpoints. Finally, the similarity sequence of the target vehicle image is obtained by the metric calculation of the sample features extracted by the trained model. Through experiments on the new constructed large-scale UAV-captured dataset called VeRi-UAV, the VD-CML algorithm has better performance on the vehicle re-identification problem compared with other methods.