`MEDJOOL` FRUIT-CLUSTER RECOGNITION FOR MOTION PLANNING

The Medjool-date fruit grow in fruit-clusters which are made up of an inflorescence rachis and spikelets. Medjool-date fruitlets require thinning for producing high quality fruit over time. Currently thinning is done manually and automation is required for enhancing work rate and work uniformity. Using automation thinning can be conducted by shortening the spikelets. A critical task for a robotic thinning system is fruit-cluster recognition and feature extraction. This is required both for deciding where to cut the spikelets and for guiding the robotic motion. The recognition is challenging since the environment is dynamic with varying lighting conditions, there are variations in cluster shape, and there are obstructions due to other plant parts.

We present an innovative algorithm for Medjool fruit-cluster identification facilitating motion planning in a robotic thinning application. The algorithm is based on image segmentation using RGB and HSV representations and morphological operations. The algorithm finds intersection between the inflorescence rachis and the spikelets based on fruit-cluster geometry, and then calculates the spikelets` oriented bounding-box. The algorithm quantifies the parameters required for the motion planning (the intersection point between the inflorescence rachis and spikelets, and the length, width, and orientation of the spikelets` bounding box). These parameters serve as input to a motion planning based on parametrized dynamic motion primitives where the parameters are learned using advanced machine learning methods.

The image processing algorithm was tested in a laboratory setup with regular florescent lighting. An apparatus was constructed to hold a Medjool date fruit-cluster and leafs, simulating cluster positioning in the date palm crown. The apparatus enables placing the cluster at different heights and angels. Nine Medjool fruit-clusters were each positioned behind two leafs at three different angels with respect to the ground (12°,20°,25°). Images were acquired with an RGB camera at a distance of 1500 mm. The algorithm was applied to each image and results were compared to ground truth data established by manually marking pixels in each image.

Correct separation between the inflorescence rachis and spikelets is a critical aspect of the algorithm, therefore images for which the separation point identification error was higher than a predefined threshold (half average length and half average width of the inflorescence rachis) were not further processed. The algorithm`s quality was determined based on the mean absolute percentage error (MAPE) between the ground truth and the observed value of each measured parameter.

The intersection point`s error was lower than the threshold for 78%, 56% and 33% of the images at 12°, 20° and 25° respectively. The mean MAPE over all parameters was 10.4%. The results show that the separation between the inflorescence rachis and spikelets is highly effected by the angle of the fruit-cluster with respect to the ground and requires an improved algorithm. Once the separation is established parameters are readily quantified. Future work will enhance the algorithm to facilitate improved Medjool fruit-cluster recognition and feature extraction in the orchard.

Acknowledgments: Research is supported by the Israeli ministry of agriculture and the Israeli Date Grower`s board in The Plant Council.