Multi-Type Target Detection and Tracking in UAV Images under Typical Scenarios

Despite significant advances in UAV target detection and tracking algorithms, real-world flight tests and data validation remain insufficient. In this study, we developed a stable and reliable UAV image target detection and tracking system based on the YOLOv5 and TCTrack networks. The system demonstrated excellent stability and reliability in laboratory-simulated scenarios, particularly excelling in target recognition accuracy. Additionally, the system was tested in real UAV flight scenarios, showcasing robust real-time performance suitable for various task scenarios. Furthermore, we collected and annotated a comprehensive outdoor multi-scene target detection dataset by operating different types of UAVs to further validate and optimize the algorithm.

I. Introduction

In recent developments, despite significant progress in algorithms for target detection and tracking in unmanned aerial vehicles (UAVs), actual flight testing and data validation remain relatively scarce. This study developed a UAV image target detection and tracking system tailored for typical scenarios, based on advanced YOLOv5[1] and TCTrack[2] networks. The system demonstrated exceptional stability and reliability across various laboratory-simulated scenarios, particularly excelling in target recognition accuracy. Additionally, by operating different types of UAVs and collecting image data, we compiled and annotated a comprehensive dataset. This research also tested the system's performance in actual flight conditions, showing that the algorithm could adapt to various task scenarios and exhibit robust real-time performance.