Vision-Based 3D Peach Tree Reconstruction for Automated Blossom Thinning

This paper presents research using a correlation-based stereo vision approach to 3D blossom mapping for automated thinning of peach blossoms on perpendicular "V" architecture trees. To this end, a calibrated camera system is proposed, based upon three synchronized ten megapixel cameras and flash illumination for nighttime image acquisition. A correlation-based stereo algorithm, suitable for parallel processing, is developed with the actual scene structure in mind using multiple camera pairs for validating 3D locations, three different certainty metrics, and does not require extrinsic rectification of the images. Results show that mapping accuracy of less than half of a blossom width (similar to 1 cm) is feasible, and validates the approach as the sensor part of an automated selective blossom thinning system. Furthermore, the effects of the different certainty metrics are examined. They effectively improve the accuracy of blossom positions when the visibility of blossoms is good by removing insecure matches and through qualified selection of subsets of cameras for 3D triangulation. The proposed algorithm is compared and found superior to a popular global optimization algorithm, designed to perform well in another scene structure, demonstrating the quality of correlation-based stereo in practical applications.