Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.
Detection of cherry tree branches for automated shake-and-catch harvesting
Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.orgCitation: 2015 ASABE Annual International Meeting 152158372.(doi:10.13031/aim.20152158372)
Authors: Suraj Amatya, Manoj Karkee
Keywords: Branch detection, cherries, automated harvesting, image processing, Bayesian classification
Abstract. To minimize the demand of seasonal workforce, mechanical shaking techniques have been widely investigated for harvesting nuts and tree fruit crops including sweet cherries. However, operating mechanical harvesters equipped with shaking mechanism is challenging for the operator to locate tree branches and position the actuator for efficient harvesting. A machine vision-based automated harvesting system could eliminate manual handling, positioning and operation of the harvester. In this study, a branch detection algorithm was developed to identify branches of cherry trees trained in Y-trellis architecture. First, Bayesian classification was used to classify images into branches, cherries, leaves and background. Branch and cherry pixels were classified with accuracies of 89.6% and 73.3% respectively. Branch and cherry regions were then used to detect tree branches. First, partially visible branch segments within the canopies were connected using morphological features of the segments and then represented by a linear or logarithmic equation. On the second part, the positions of cherry clusters in the canopy were used as an indication to detect branches that were completely occluded by cherries and leaves. This second method of branch detection was possible because the locations of cherry clusters are generally in close proximity to a branch. Different cherry clusters were grouped together based on their spatial location and distance between them. Branch equations were defined through those cherry clusters using minimum residual criteria. In total, 92.1% branches were detected in the Y-trellis fruiting wall cherry canopy, with 60.4% of branches detected using only branch pixels and 31.7% additional branches detected using cherry clusters. The method resulted in a total of 11.3% of false positive detection. The results showed that branch detection accuracy can be substantially improved by integrating cherry location information with the location of segments of partially visible branches. This study has shown the potential for the application of machine vision system to detect cherry tree branches in full foliage season, which is highly promising for the development of automated sweet cherry harvesting systems.(Download PDF) (Export to EndNotes)