Abstract:

At present, the existing unmanned aerial vehicle （UAV） tracking methods have problems such as low detection accuracy for long-range UAVs, large parameter quantities that are difficult to track in real time, and easy target loss. Therefore, a lightweight UAV tracking method based on improved you only look once version 8（YOLOv8） is proposed. In response to the problem of low detection accuracy of long-range UAVs using existing methods, YOLOv8 is used as the baseline model to replace the original convolution module in the network structure with spatial to deep grouped convolution, which improves the model’s feature extraction ability for small targets while reducing network parameters. To address the problem of difficulty in real-time tracking due to the large number of model parameters, a deep separable shuffle network structure is designed as the backbone network of the model, which reduces the number of model parameters while ensuring detection accuracy. To address the issue of tracking loss in ordinary tracking models, an improved detection model combined with ByteTrack algorithm is used to enhance the tracking performance of UAVs in complex environments. The tracking method is validated on the Real World dataset, and compared to the baseline model, the improved UAV detection model shows a 1.6% increase in detection accuracy, a 0.8% increase in recall, a 0.2 increase in F1 metric value, a 0.5% increase in average detection accuracy, and a 0.2×10⁶ reduction in parameter count, demonstrating that the model has good detection accuracy and real-time performance. Tracking test is conducted on UAV flight videos, and the results show that the proposed method has good performance in UAV tracking.

Key words: deep learning, target detection, target tracking, deep separable network structure, you only look once (YOLO)