TFAtrack: Temporal Feature Aggregation for UAV Tracking and a Unified Benchmark

Xiaowei Zhao, Youhua Zhang

Abstract: Research on object detection and tracking has achieved remarkable progress in recent years. Due to the superior viewing angle and maneuverability advantages of unmanned aerial vehicles (UAVs), the application of UAV-based tracking is also undergoing rapid development. But since the targets captured by UAVs are tiny and all have similarities and low recognition, this leads to the great challenge of multiple-object tracking (MOT). To solve the two problems mentioned above, We propose TFATracking, a comprehensive framework that fully exploits temporal context for UAV tracking. To further reflect the effectiveness of the algorithm and promote the development of UAV object tracking, we present a large-scale, high-diversity benchmark for short-term UAV multi-object tracking named T2UAV in this work. It contains 20 UAV-captured video sequences with a total number of frames over 12k and an average video length of over 600 frames. We conduct a comprehensive performance evaluation of 8 MOT algorithms on the dataset and present a detailed analysis. We will release the dataset for free academic use.

External IDs:dblp:conf/prcv/ZhaoZ22