For a fruit detection system of a tomato harvesting robot, a background removal algorithm was developed to separate foreground and background plants using a depth camera. This algorithm enabled the detection of only foreground fruit and improved the accuracy of the detection system. The background removal algorithm could extract only the fruits, leaves, and stems in the foreground. The F-measure was 0.894 after examining the overlap between the images generated using the background removal algorithm and the correct images. The results of the learning experiments using Faster R-CNN, SSD, and YOLO with the background removal images showed that the F-measure improved from 0.965 to 0.987 in YOLO, demonstrating the achievement of background-independent learning with reduced training data creation costs.

A tractor-mounted multi-camera system was developed and compared with the conventionally utilised UAV-RGB for phenotyping purposes. The system consisted of 6 cameras arrayed in oblique and nadir angles thus enabling sufficient overlap of images to be obtained even when the tractor was in motion. By utilising an automatic image processing pipeline (PREPs), the height, volume and canopy coverage change of three potato crop varieties: Kitahime, Touya and Toyoshiro, and the effect of compost manure treatment was estimated. In both systems, Toyoshiro had the highest volume during the peak of the growth season. A high linear correlation with a coefficient of determination (R²) of 0.97 was obtained between the crop height estimated by the multi-camera system and that estimated by the UAV.