Camera-Trajectory-Driven Signed Distance Fields for Automated Object Isolation
12th International Conference on Automation, Robotics and Applications, ICARA 2026, İstanbul, Türkiye, 5 - 07 Şubat 2026, ss.134-139, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/icara69401.2026.11480350
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.134-139
- Anahtar Kelimeler: drone trajectory, gaussian splatting, pose reconstruction, signed distance field
- Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet
Özet
Accurate isolated object reconstruction from drone footage is often challenging because of background clutter, sparse camera coverage, and viewpoint limitations. We introduce a geometric method in this study that leverages the drone's camera positions and viewing directions, extracted from Structure-fromMotion, to construct a kernel-based Signed Distance Field. This field explicitly encodes whether points lie inside, on, or outside the drone's flight route, providing a precise spatial representation of the trajectory. By evaluating the SDF at candidate points, the method can identify regions corresponding to the object of interest versus the surrounding environment. In typical SfM pipelines such as COLMAP, unbounded scenes or spiral drone captures often lead to sparse reconstructions that unintentionally include distant background structures, which in turn causes Gaussian Splats to represent unwanted terrain or sky regions. Our trajectory-aligned SDF mitigates this issue by enabling posttraining filtering of Gaussian Splat reconstructions, removing background splats (e.g., terrain/sky/background buildings) without changing the original splatting pipeline. We validate the approach on synthetic spiral trajectories that include wind and sensor disturbances, as well as on real drone flights. Results demonstrate that the field generates smooth, consistent spatial representations, offering a foundation for trajectory-informed 3-Dimensional reconstruction, automated object isolation, and targeted scene editing. This method provides a practical and flexible framework for future drone-based data capture and geometric analysis.