Abstract:
Maintaining the massive forests of the forestry industry requires physical samples from the tree canopy to enable breeding, genetics research and to monitor for diseases. The limitations of existing canopy sampling methods, particularly their restrictive nature and health and safety risks, led to the development of an autonomous UAV canopy sampling solution. Fundamentally, successful branch interception relies on knowing its relative location and trajectory to within the maximum gripper opening of 115mm. This requires locating and tracking a highly unstructured, moving object against a dynamic, cluttered background to enable outdoor canopy sampling. The developed algorithm utilises stereo vision techniques to reconstruct a 3D point cloud of the target branch from which the intercept point location is subsequently estimated. The optimal parameters were selected by quantifying the accuracy against ground truth obtained from motion capture. For a stationary branch with a plain background, the intercept point error is estimated to be 21:5mm, 40:4mm and 111mm at distances of 1m, 2m and 3m from the camera respectively. Since wind will likely affect the tree canopy, the branch will exhibit some sway motion which can be modelled as simple harmonic motion. A linear Kalman filter and a non-linear extended Kalman filter were formulated to estimate the sway parameters. The former models the intercept point motion as the linear translation of a point mass while the latter models the branch motion as rigid body rotation. The presence of motion blur and pine needle motion led to higher intercept point errors. At a distance of 1m with a plain background, the error across a variety of camera perspectives is estimated to be 35:8mm, 42:1mm and 45:8mm for sway amplitudes of 50mm, 150mm and 250mm respectively. When tested with a cluttered background, the reduced contrast resulted in higher standard deviations. Overall, the simpler model of the Kalman filter led to higher reductions in the standard deviation, averaging 17% and 33% for plain and cluttered backgrounds respectively. Subsequent outdoor testing verified the capability of the developed algorithm to estimate the branch trajectory to well within the maximum gripper opening.