Detection of cherry tree branches and localization of shaking positions for automated sweet cherry harvesting

Thesis (Ph.D.), Department of Biological Systems Engineering, Washington State UniversitySweet cherry industry has been challenged by increasingly uncertain availability and rising cost of farm labor. As harvesting is the most labor-intensive operation, there is a need for developing automated harvesting solutions to address labor related challenges. Over the past several decades, researchers have studied mechanical shakers for efficient sweet cherry harvesting. Automating a shaking harvester using a machine vision system for detecting and locating tree branches has the potential to further reduce the labor demand for cherry harvesting. This research focused on detecting cherry tree branches in full foliage canopies. The branch detection method used visible branch sections in canopy images to reconstruct whole branches. Morphological features of the visible branch sections including orientation, length and thickness were used to group sections of the same branch together, which were then connected using a model equation. This method achieved a branch detection accuracy of 89% in cherry trees trained in vertical trellis system. Using this method, in Y-trellis system with denser foliage, only 55% branches were detected. Therefore, cherry clusters were detected and integrated for locating highly occluded branches. This method identified series of cherry clusters growing along specific branches and a model equation was fitted to reconstruct complete branches. This method led to a branch detection accuracy of 94% in Y-trellis system.After branch detection, a method was developed to determine shaking positions on branches for effective cherry harvesting. Shaking locations were first identified in the color images. For each shaking position identified in the color image, 3D location information was estimated using 3D camera images. The distance to shaking positions from the camera was estimated with a root mean square value of 0.064 m with this method. Cherry tree branches were then shaken at the estimated locations using a handheld shaker. Maximum fruit removal efficiency was found to be 93% and 87% in Y- and vertical trellis system respectively. The results of this research have shown a great potential for machine vision-based automated solutions for cherry harvesting.Department of Biological Systems Engineering, Washington State Universit