Accuracy Analysis of Three-Dimensional Model Reconstructed by Spherical Video Images

Indoor 3D model is useful for navigation purpose and virtual reality applications. However, its surface is
usually texture-less with unbalanced illumination situations together with complex architecture design, which is difficult
to acquire images and produce the high-quality 3D model. In this study, we propose the use of spherical video images
due to its 720 degrees field-of-view (FOV) and its sequential data acquisition is friendly to the user. In order to evaluate
its performance, positioning accuracy and feasibility, we also acquire terrestrial laser scanning (TLS) data for comparison
and measure ground control points (GCPs) by a total station for accuracy analysis. The Garmin VIRB 360 spherical
camera is used to acquire video at 30 fps (frame per second) with an image size of 3840 x 2178. The video is converted
to static image sequence with a time interval of 1.23 seconds. The stitched spherical images were imported into Agisoft
Photoscan Pro for 3D modeling. The photo triangulation is conduct by Structure-from-Motion (SfM) technique and the
positioning accuracy was assessed by GCPs within the Photoscan Pro software. Later, a dense point cloud, mesh model
as well as photo-realistic texture model can be produced. The produced dense point cloud was compared with TLS data
within CloudCompare software. The experiment result shown that the 3D positioning accuracy (RMSE) after SfM is
about 18.9 cm, while the 3D distance discrepancy between the produced dense point cloud with TLS data is about 40.3
cm. The major problem that causes such a large error is the stitching algorithm to create a spherical image does not follow
the photogrammetric collinearity condition. It demonstrates that the adoption of Garmin VIRB 360 spherical camera for
3D modeling is beneficial in data acquisition, but its positioning accuracy is too low for high-accuracy applications and
may be applicable to indoor navigation applications.