A small unmanned aerial vehicle (sUAV) can be used to reconstruct a 3D scene by capturing frames consisting of LiDAR and aerial photography data, creating a textured digital surface model (TDSM) with the full LiDAR point cloud and an overlaid registered image. Forming a complete 3D scene using texel images (fused LiDAR and digital image scans) from an entire flight can be computationally prohibitive on low-cost hardware, so a streaming bundle adjustment algorithm can be used to process the data using a sliding window. The streaming algorithm uses less memory and is faster than a full bundle adjustment. Depending on the flight pattern, matching points in the scene may be visible from frames which are not adjacent in time, so reconstructing a complete scene can take into account matching points from non-adjacent frames to better correct for error.
A modification to the streaming bundle adjustment algorithm is described which finds overlap in frames which are not adjacent in time to correct for error. This algorithm also addresses the loop-closing problem that occurs when the sensor returns to the starting point of a survey. Flight data from a sensor constructed with low cost, commercial off-the-shelf parts is used to demonstrate how 3D scene reconstruction using this algorithm can correct for errors compared with data gathered from a full-scale aircraft. Examples of the resulting TDSMs are presented.
|