Título: RTD: A Video Dataset for Runway Tracking Applications
Autores: João Pedro Klock Ferreira, Cristiano Leite de Castro, João Paulo Lara, Júlia Santos Moura & Gabriel Lott
Resumo: Recent advances in the aerial industry with the advent of computer vision have led to the need for large image datasets in this field. Specifically, there is a lack of video datasets for the task of tracking the runway during a flight landing process. In this paper, we propose a new dataset with 20 videos, where we propose 10 manually annotated videos of real planes landing, and 10 synthetic videos from the Video-LARD dataset [1] with new weather conditions and lightning variations, generated through UniSVST [2], a diffusion image-to-video model. We also perform careful experiments using a pre-trained version of the LoRAT tracking model [3] to verify the quality of our dataset, assessing its strengths and weaknesses. The results show that the model obtained better tracking performance when closer to the runway for real videos, and for synthetic videos, blurry climates such as fog or dark nights led to a worse performance. Overall, the dataset incorporates a variety of difficult tracking conditions, but synthetic video generation can still be improved. The dataset and any relevant code are publicly available at https://github.com/jpklock2/RTD
Palavras-chave: object tracking; video dataset; runway; diffusion model; plane landing.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1191933
Artigo em PDF: CBIC_2025_paper1191933.pdf
Arquivo BibTeX:
CBIC_2025_1191933.bib