By Jesse Murray
Advancements in unmanned aerial vehicle (UAV) technology have made UAVs easier to operate, more affordable, and capable of performing a broader range of tasks. This trend has led to a significant expansion of UAV applications in society. However, the use of UAVs for tasks that require high precision measurements is still an area of development due to challenges in determining the precise location at which a UAV collects the desired data. While on-board, high-precision Global Navigation Satellite System (GNSS) sensors can be employed to provide such information, this adds significant cost and operational complexity to the operation of a UAV. Moreover, such GNSS sensors cannot be used in indoor or complex urban or industrial environments where the GNSS signals become unreliable or completely fail. The objective of this thesis is to produce a highly precise localization of a moving UAV using a network of static, non time-synchronized video surveillance cameras. The pipeline uses pixel-wise detection sequences of a moving UAV captured in the image space of each network camera. A robust solver is implemented to simultaneously estimate the geometry of the camera network and the relative temporal offsets between the video sequences of the moving drone. An incremental bundle adjustment procedure is used to jointly optimize the relative camera geometries and UAV trajectory. The bundle adjustment procedure integrates rolling shutter correction and several motion constraint methods including B-splines and physical motion priors. Experiments using various models of the bundle adjustment procedure were conducted on synthetic and outdoor video data to determine the most robust and precise configuration. The results of the reconstructed trajectories obtained on the outdoor data show that the implemented approach can obtain decimeter accuracy as compared to Real-time kinematic (RTK) ground truth measurements collected during the flight.