Thanks to visit codestin.com
Credit goes to github.com

Skip to content

A 3D object detection and tracking system for UAVs that predicts potential mid-air collisions. Also plan, optimise, and validate UAV missions using GPS waypoints and coordinate transformations across WGS84, ECEF, NED, and FRD frames

Notifications You must be signed in to change notification settings

afnarimzi/kestrel

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

26 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Kestrel

This project develops a proof-of-concept for an onboard, real-time computer vision system designed to detect birds, track their 3D movement, and predict potential mid-air collisions with UAVs and validate UAV missions using GPS waypoints and coordinate transformations across WGS84, ECEF, NED, and FRD frames

Methodology and Innovation

The system is engineered in three integrated stages, solving key technical challenges:

1. Data Solution: Synthetic Training

Challenge: The extreme scarcity of real-world cockpit footage showing birds in various conditions.

Solution: Solved the problem by leveraging Generative AI (Stable Diffusion) to synthesize a custom, robust dataset of 324 aerial bird images for model training. The dataset was meticulously labeled using manual Data Labelling (Bounding Boxes).

2. 3D Localization and Tracking

Challenge: Converting the flat, 2D pixel coordinates from a camera into the bird's true position and velocity in 3D space.

Solution: Integrated Stereo Triangulation and Spatial Math (using SE3 transforms) to convert simulated two-camera detections into precise 3D position vectors and calculate the bird's subsequent velocity.

3. Real-Time Threat Assessment

Challenge: Determining if the calculated 3D path of the bird intersects the aircraft's path.

Solution: Implemented a path projection algorithm that uses the calculated bird velocity and the aircraft's simulated flight vector to compute the predicted Closest Point of Approach (CPA) and trigger a COLLISION ALERT if the distance falls below the safety threshold.

Project Deliverables & Results

1. Model Performance

Custom Model Trained: A custom best.pt YOLOv8 model was successfully fine-tuned on the GPU using the synthetic dataset.

Accuracy: The model achieved high accuracy metrics (e.g., mAP50 > 0.99) on the validation set, validating the quality of the synthetic data.

2. Final Pipeline Output

The final_pipeline.py script successfully ran the end-to-end simulation:

System Validation: The pipeline consistently detects the simulated bird, calculates its path, and triggers a verifiable COLLISION ALERT when the prediction logic dictates a threat below the 100-meter safety threshold.

Actionable Metric: The system outputs the exact Time-to-Impact (TTI), providing the essential warning metric for a autonomous system.

About

A 3D object detection and tracking system for UAVs that predicts potential mid-air collisions. Also plan, optimise, and validate UAV missions using GPS waypoints and coordinate transformations across WGS84, ECEF, NED, and FRD frames

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages