Drone-to-Drone Tracking

• Constructed datasets for model training. I annotated drone-to-drone tracking datasets which are provided by project client. And added self-synthesized data, web-scrape data, also used the data generated by the stable diffusion model to increase the amount of data. The drone dataset contained various conditions including mountains, forests, buildings, rainy, foggy, twilight and so on.
• Used YOLOv8 model as the detection model and the deep-sort model as the tracking algorithm to implement drone tracking. After training the YOLOv8 model using a large amount of training data, the model can obtain good tracking results. Tuned the model in order to have a good performance on edge computing devices.
• Constructed the tracking model by splicing the CSP DarkNet53 backbone network and the Spatio-Temporal Swin Transformer. This design allows for efficient extraction of spatial and temporal features from the input frames, enabling robust tracking performance. Even in scenarios where the target drone flying into complex backgrounds, the model can still achieve good tracking.
• Deployed the tracking models with good performance to the Raspberry Pi chip of the drone to achieve real-time tracking applications. Evaluated and iterated the models, and tested the tracking effect, metrics, and FPS of the models.