A stereo dataset of annotated budgerigar flight trajectories for multi-agent collision avoidance studies

Area of Science:

• Animal behavior and biomechanics
• Computer vision and machine learning
• Robotics and autonomous systems

Background:

• Understanding avian flight dynamics is crucial for bio-inspired robotics and aerodynamic research.
• Existing datasets often lack synchronized 3D motion capture or detailed behavioral annotations.
• Budgerigars offer a tractable model for studying complex flight maneuvers and social interactions.

Purpose of the Study:

• To present a comprehensive dataset of budgerigar flight behavior.
• To enable advanced analysis of flight dynamics, trajectory, and social interactions in birds.
• To provide a resource for developing and validating algorithms in multi-object tracking and 3D pose estimation.

Main Methods:

• Acquisition of 360 synchronized stereo video pairs of budgerigar flight at 120 fps.
• Manual annotation of 2,760,178 identity-consistent 2D bounding boxes and keypoints (head, tail, wings).
• Reconstruction of 3D trajectories and poses using calibrated stereo triangulation.

Main Results:

• A dataset containing synchronized stereo videos, 2D/3D annotations, and calibration data.
• Technical validation including stereo reprojection error (0.50 pixels) and motion plausibility.
• Dataset supports applications in multi-object tracking, 3D pose estimation, and multi-agent interaction analysis.

Conclusions:

• The budgerigar flight dataset provides a valuable resource for studying avian locomotion and behavior.
• The integrated 2D and 3D data facilitates research in biomechanics, computer vision, and AI.
• Future work could explore wing kinematics, though limited by current data acquisition constraints.