An Experimental Evaluation Based on New Air-to-Air Multi-UAV Tracking Dataset

Visual-based multi-object tracking (MOT) of micro unmanned aerial vehicles (UAV s) is a crucial technology that plays a significant role in advancing the development of UAV s. It can be applied in cooperative UAV formation, UAV countermeasure systems, multi-UAV logistics and other fields. However, the performance of existing visual-based MOT algorithms in UAVs has yet to be evaluated. To alleviate this situation, we provide a comprehensive air-to-air multi-UAV tracking dataset, MOT-FLY, which includes more than 11 000 images of three types of UAVs. The dataset encompasses various backgrounds, viewing angles, lighting conditions, object sizes, target movement patterns, and challenging scenarios. Additionally, this paper designs evaluation experiments on eight representative MOT algorithms using the proposed dataset. The results indicate that dataset composition, network structure, and image characteristics all have an impact on the algorithm's performance. Based on these findings, we provide recommendations to address the challenges faced by air-to-air multi-UAV tracking algorithms. The MOT-FLY dataset is published at https://github.com/CZC-123IMOT-FLY.