Deep learning for unmanned aerial vehicle re-identification

Abstract

Recently, Unmanned Aerial Vehicle (UAV) Re-Identification (ReID) has gained significant popularity in computer vision and ReID. This emerging technology holds considerable significance in urban security and military applications. In contrast to traditional ReID methods that depend on stationary cameras with fixed perspectives, UAV-based ReID encounters more challenging obstacles. UAVs’ flexible and movable viewpoints introduce more appearance ambiguities among objects, making ReID more challenging. Moreover, the computing and power limitation of UAV platforms poses constraints on the use of large-scale deep learning networks. The above factors result in the inherent difficulty in achieving high accuracy in UAV-based ReID research. To solve these problems, this dissertation employs three baselines from FastReID to conduct experiments in two scenarios: UAV-based person ReID and vehicle ReID. After comparison, the Stronger Baseline (SBS) model with ResNeSt as backbone network and Instance and Batch Normalization (IBN) performed best for UAV-based person and vehicle ReID, which reduced processing time and computing costs. Through detailed and comprehensive experiments and in-depth analysis of the results, valuable insights are provided for future studies on UAV-based ReID.