A Deep Learning Approach for Unifying Object Re-Identification and Cross-view Geo-localization on Autonomous UAVs

In recent years, with the rapid development of Unmanned Autonomous Vehicles (UAVs) technologies, the applications of autonomous UAVs in aerial surveillance-related missions have been increasing. These missions often involve image processing-related tasks, such as object (human, vehicle) detection, re-identification, image-based localization, and navigation. However, these tasks often require much computation to achieve the best performance, which is quite a challenge for autonomous UAVs due to the limitation of hardware and computing resources on these devices. Unfortunately, no architecture has been established to perform these tasks adequately. In this study, for the first time, we combine three tasks that can significantly contribute to the operation of autonomous UAVs in real-life missions: Person Re-Identification, Vehicle Re-Identification, and Cross-view Geo-Localization. We exploit the generalization ability of currently available deep metric learning methods and propose a framework to learn to re-identify objects (humans, vehicles) and perform geo-localization simultaneously. The proposed method was evaluated on benchmark UAV datasets and received promising results.