UAV target detection and tracking technology based on deep learning algorithms

Nowadays, there is an increasing number of tasks on unmanned aerial vehicles (UAVs), which require accurate and real-time analysis ability. In this work, we propose a lightweight and efficient model for object detection and tracking in UAV videos. We use Tiny-YOLOv5 as the detector while DeepSORT is implemented as the tracking module. This combination provides high quality results with low computational cost. Meanwhile, we also introduce multi-modal inputs to improve the detection performance in challenging environments. These inputs are fused with a simple early operation before being passed to the detection model. We evaluate our method on several public UAV datasets. As is shown in results, our system achieves better detection accuracy and tracking stability than baseline models. Meanwhile, according to the evaluation, our model still remains a low computational cost suitable for real-time UAV applications.