Markerless deformation capture of hoverfly wings using multiple calibrated cameras

Abstract

This thesis introduces an algorithm for the automated deformation capture of hoverfly wings from multiple camera image sequences. The algorithm is capable of extracting dense surface measurements, without the aid of fiducial markers, over an arbitrary number of wingbeats of hovering flight and requires limited manual initialisation. A novel motion prediction method, called the ‘normalised stroke model’, makes use of the similarity of adjacent wing strokes to predict wing keypoint locations, which are then iteratively refined in a stereo image registration procedure. Outlier removal, wing fitting and further refinement using independently reconstructed boundary points complete the algorithm. It was tested on two hovering data sets, as well as a challenging flight manoeuvre. By comparing the 3-d positions of keypoints extracted from these surfaces with those resulting from manual identification, the accuracy of the algorithm is shown to approach that of a fully manual approach. In particular, half of the algorithm-extracted keypoints were within 0.17mm of manually identified keypoints, approximately equal to the error of the manual identification process. This algorithm is unique among purely image based flapping flight studies in the level of automation it achieves, and its generality would make it applicable to wing tracking of other insects.