Transformer-Based Multi-object Tracking in Unmanned Aerial Vehicles

Unmanned aerial vehicles (UAVs) possess a wide field of view, maneuverability, and autonomy in the field of multi-object tracking. By combining the high-altitude perspective and multidimensional perception capabilities of UAVs with artificial intelligence image processing and object recognition technologies, efficient and accurate multi-object tracking can be achieved. To enhance the processing capacity of relevant information, a transformer-based multi-object tracking model is proposed in UAV perspectives. The backbone network of detector is based on the transformer architecture to extract target features. In this structure, the network is further deepened to optimize feature fusion and improve the ability to capture small targets. To address the tracking challenges arising from dynamic variations between different targets, the tracker implements multi-object tracking in UAV scenarios based on the motion characteristics of targets and optimizes the tracking process accordingly. The model was evaluated on the Visdrone dataset, achieving a detection accuracy of 90.1% and a MOTA of 51.9%. Furthermore, the model demonstrated a remarkable speed of 20 frames per second. In summary, the proposed algorithm demonstrates excellent performance in target tracking from UAV perspectives and exhibits certain advantages over other algorithms.