Objectives Incremental structure from motion (SfM) has become the widely used workflow for aerial triangulation of unmanned aerial vehicle (UAV) images. However, it suffers from significant issues of error accumulation and efficiency degradation. The existing parallelization methods for SfM suffer from issues such as the scene graph construction ignoring the local connection strength of images, long merging paths for sub-scenes (spatially related image collections), and low efficiency, making it difficult to meet the requirements of large-scale scene image processing.

Methods This paper proposes a parallelized SfM method for UAV images that does not rely on fixed anchor points (reference models). The proposed solution develops a scene graph construction algorithm that considers local connection strength constraints, effectively improving the compactness of scene graph partitioning and ensuring the high robustness of SfM reconstruction within partitions. Then, it proposes a parallelized sub-scene merging method that does not rely on fixed anchor points, enabling adaptive adjustment of merging paths and parallel merging of multiple sub-scenes, significantly enhancing the reliability and efficiency of sub-scene merging. To avoid the influence of outliers in corresponding geometric elements, this paper employs the random sample consensus hypothesis-verification framework for robust parameter estimation and designs a transformation model estimation based on a bidirectional mean square reprojection error metric.

Results Experiments using real UAV images demonstrate that the proposed method achieves a 9 times speedup in sub-scene merging, with relative and absolute orientation performance comparable to the existing methods. And the proposed method has also been successfully applied to the parallel SfM reconstruction of nearly 90 000 UAV images covering an area of about 50 km².

Conclusions Considering the metric of efficiency and precision, the proposed method can be a validate solution for the efficient SfM reconstruction for large-scale UAV datasets.