Aero1track3D

Inspiration

                          Disaster response, inspection and survey teams usually need several drone flights and hours of photogrammetry before they get a usable 3D map. Cloud processing also fails in the field, where connectivity is poor. We wanted a usable map from one pass, processed near the edge, with honest reporting of how reliable it is.

What it does

Aero1Track3D takes drone video, GPS data, frame metadata and optional RTK/IMU data. It produces a textured, georeferenced 3D model in an interactive viewer. Users can: *Measure distances, heights, areas and elevation. *Validate the 3D geometry. *Ask questions through “Twin AI”, a lightweight local natural-language query layer.

A “Trust 3D” panel reports reconstruction quality, feature matching, georeferencing and coverage, plus an overall confidence score. It separates reconstruction confidence from true spatial accuracy, so sensor data is reported and never assumed.

How we built it

  The pipeline start with 
                   1)Uploading the input source given , 
                    2)Video decoding.
                    3)Quality filtering and frame selection (“Frame Intelligence”).
                                               OpenCV: decodes frames (cv2.VideoCapture), then scores each frame for blur and exposure (Laplacian variance, histogram checks) so only sharp, well-exposed frames go forward.
                    4)Feature extraction and matching with SIFT/ALIKED.
                    5)Structure-from-Motion for camera poses and a sparse cloud.
                    6)Multi-View Stereo and meshing for a dense model.
                    7)WGS84 GPS alignment, with RTK/IMU data to improve geolocation.
                    8)Model optimization and the Trust 3D quality check.
                    9)Export to GLB/PLY with an interactive viewer, plus spatial analysis and Twin AI queries.

                   10) A YOLO26n detector fine-tuned on aerial data handles dynamic-object detection.

Challenges we ran into

                 1)Single viewing angle, motion blur, changing light and dynamic objects.
                 2)GPS noise, occlusion and speed.
                 3)Getting accuracy without ground control points (GCPs).

Our answer was frame-quality filtering, RTK-assisted georeferencing and the Trust 3D score.

Accomplishments that we're proud of

                 A working prototype flow, with UI screenshots, a prototype report and a demo video.

Reported results: under 14 minutes to process a 10-minute video, about 97% frame reduction, measurement accuracy under 8.3 cm, 95% detection sensitivity, a 27% processing-time reduction versus photogrammetry, and +5% detection accuracy from fine-tuning.

What we learned

                   We learned that OpenCV-based frame selection, such as blur and exposure scoring, determines which frames reach feature matching and 3D reconstruction. We also learned why a reconstruction needs an uncertainty report such as Trust 3D as much as an accuracy figure, and how GNSS/RTK data can be combined with vision-based 3D reconstruction.

What's next for Aero1track3D

                     1)Onboard edge deployment.
                     2)Better coverage through multi-angle capture.
                     3)Tests against ground-control benchmarks.
           Use cases: disaster assessment, infrastructure inspection, construction monitoring, smart-city mapping.