Probabilistic Bird Trajectory Forecasting with Heavy-Tailed Uncertainty Modeling for Low-Altitude Airspace Monitoring

Area of Science:

• Computer Vision
• Robotics
• Aerospace Engineering

Background:

• Low-altitude airspace is increasingly shared by birds and unmanned aerial vehicles (UAVs), creating safety concerns.
• Existing trajectory prediction methods often fail to account for heavy-tailed flight dynamics and lack uncertainty calibration.
• Current vision-based systems treat detection and prediction as separate tasks, limiting real-world deployment.

Purpose of the Study:

• To develop a unified framework for low-altitude airspace monitoring, integrating trajectory forecasting and UAV detection.
• To address limitations of existing methods by modeling heavy-tailed flight dynamics and providing calibrated uncertainty.
• To enable efficient, deployable solutions for shared airspace safety.

Main Methods:

• Proposed Mini-BirdFormer, a lightweight Transformer encoder combined with a Student-t mixture density head.
• Modeled heavy-tailed flight dynamics for accurate bird and UAV trajectory prediction.
• Integrated a system-level extension for zero-shot UAV detection using open-vocabulary learning.

Main Results:

• Achieved strong long-horizon trajectory forecasting performance (minADE of 0.785 m) with only 1.05 million parameters.
• Significantly improved uncertainty calibration, reducing negative log-likelihood from 1.25 to -2.01 compared to a Gaussian LSTM baseline.
• Enabled low-latency inference at 616 FPS on resource-constrained platforms and attained 92% recall for zero-shot UAV detection.

Conclusions:

• The proposed framework offers a practical and deployable solution for monitoring shared low-altitude airspace.
• Combining heavy-tailed probabilistic modeling with a compact backbone enhances trajectory forecasting and uncertainty estimation.
• The system effectively addresses safety risks posed by the increasing presence of UAVs in bird-inhabited airspace.