Distant Object Localisation from Noisy Image Segmentation Sequences

Julius Pesonen, Arno Solin, Eija Honkavaara

Abstract

3D object localisation based on a sequence of camera measurements is essential for safety-critical surveillance tasks, such as drone-based wildfire monitoring. Localisation of objects detected with a camera can typically be solved with specialised sensor configurations or 3D scene reconstruction. However, in the context of distant objects or tasks limited by the amount of available computational resources, neither solution is feasible. In this paper, we show that the task can be solved with either multi-view triangulation or particle filters, with the latter also providing shape and uncertainty estimates. We studied the solutions using 3D simulation and drone-based image segmentation sequences with global navigation satellite system (GNSS) based camera pose estimates. The results suggest that combining the proposed methods with pre-existing image segmentation models and drone-carried computational resources yields a reliable system for drone-based wildfire monitoring. The proposed solutions are independent of the detection method, also enabling quick adaptation to similar tasks.

Introduction

Localizing distant hazards like wildfires from a moving Unmanned Aerial Vehicle (UAV) presents significant challenges for conventional computer vision. While 3D localization is well-established for proximal objects, accuracy degrades significantly at distances of several kilometers. Standard solutions like stereo vision require impractical baselines for small drones, while Lidar resolution decreases cubically with distance. Furthermore, full 3D reconstruction techniques such as Structure from Motion (SfM) or SLAM are often too computationally intensive for real-time onboard processing in remote areas. This paper explores lightweight probabilistic methods—specifically multi-view triangulation and particle filters—to achieve robust localization using standard RGB cameras and GNSS data.

Methodology

Two primary mathematical frameworks were evaluated for the distant target localization problem:

  1. Multi-view Triangulation (with RANSAC): This method treats detections from different drone positions as rays in 3D space, identifying the target at their intersection. To handle "outliers" caused by erroneous detections, the Random Sample Consensus (RANSAC) algorithm is employed to find a robust consensus among data points.
  2. Particle Filters: This probabilistic approach maintains a distribution of thousands of "particles" representing hypothetical target locations. As the UAV moves, particles are updated and resampled based on visual alignment.

Unlike point-based methods, the particle filter provides a probability "cloud," enabling the estimation of both the target's physical shape (e.g., smoke plumes) and the measurement uncertainty, which is typically elongated along the line of sight due to depth uncertainty.

Conceptual overview of drone-based wildfire monitoring and target localization.

Experimental Evaluation

The proposed methods were validated using a combination of synthetic and real-world data. A custom 3D simulation environment allowed for precise control over noise levels, including pose errors and false detections. Real-world testing involved drone sequences of a stationary telecom mast (700m) and industrial smoke from a chimney (1,770m). For the latter, the Segment Anything Model (SAM) was utilized to generate initial image masks for the localization algorithms.

Results and Discussion

Simulations demonstrated that the models could achieve a relative Root Mean Square Error (RMSE) of less than 5% of the target distance after the UAV traveled just a few hundred meters. In the real-world smoke tracking scenario (1.7km distance), estimates stabilized after 150–180 meters of flight, with a final RMSE between 200 and 350 meters.

While RANSAC-based triangulation was faster at identifying a central coordinate, the particle filter proved more robust to intermittent false negatives (where the target is temporarily obscured). Furthermore, the particle filter's ability to model the target's spatial extent makes it highly functional for wildfire management, where identifying the correct forest block is the primary objective.

Simulation results for multi-target localization showing segmentation masks and RMSE convergence over camera translation.

Conclusion

This study confirms that lightweight probabilistic methods can effectively localize distant objects using edge-computing hardware like the NVIDIA Jetson Orin NX, bypassing the need for high-bandwidth cloud connectivity. Future work will focus on extending these methods to handle multi-target scenarios where spot fires may appear, disappear, or merge, and testing under more diverse environmental conditions.

Summary updated

Abstract

3D object localisation based on a sequence of camera measurements is essential for safety-critical surveillance tasks, such as drone-based wildfire monitoring. Localisation of objects detected with a camera can typically be solved with specialised sensor configurations or 3D scene reconstruction. However, in the context of distant objects or tasks limited by the amount of available computational resources, neither solution is feasible. In this paper, we show that the task can be solved with either multi-view triangulation or particle filters, with the latter also providing shape and uncertainty estimates. We studied the solutions using 3D simulation and drone-based image segmentation sequences with global navigation satellite system (GNSS) based camera pose estimates. The results suggest that combining the proposed methods with pre-existing image segmentation models and drone-carried computational resources yields a reliable system for drone-based wildfire monitoring. The proposed solutions are independent of the detection method, also enabling quick adaptation to similar tasks.
BibTeX
@misc{pesonen2025distant,
  author = {Julius Pesonen and Arno Solin and Eija Honkavaara},
  title = {{Distant Object Localisation from Noisy Image Segmentation Sequences}},
  year = {2025},
  eprint = {2509.20906},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.20906},
}
Figure