An Alternative Cost Function to Bundle Adjustment Used for Aerial Photography from Uavs

The Autonomous Vehicle Systems Lab at Kansas State University has developed an unmanned aerial vehicle capability for inexpensively performing aerial surveys. Bundle Adjustment (BA) is used in the extraction of terrain information from these aerial surveys. However, the inexpensive, lightweight inertial navigation system on the UAV does not give good initial estimates of the camera poses for the BA algorithm. This leads to an incorrect convergence of the BA minimization in some cases. This paper presents an alternate formulation of BA that uses a reduced set of parameters that are varied in the minimization. For photographic sets covering large areas, but having overlap only between adjacent photos, the search space is significantly reduced when compared to standard BA. The usefulness of the algorithm is demonstrated by generating a digital elevation model of a region covered by 11 photos and an ortho-rectified composite image of this region. While this alternate formulation provides a reduced search space and significant insight about BA itself, the standard formulation seems to provide similar convergence properties in a direct comparison.