Determination of the number of green apples in RGB images recorded in orchards

This work details the development and validation of an algorithm for estimating the number of apples in color images acquired in orchards under natural illumination. Ultimately, this algorithm is intended to enable estimation of the orchard yield and be part of a management decision support system. The algorithm includes four main steps: detection of pixels that have a high probability of belonging to apples, using color and smoothness; formation and extension of "seed areas", which are connected sets of pixels that have a high probability of belonging to apples; segmentation of the contours of these seed areas into arcs and amorphous segments; and combination of these arcs and comparison of the resulting circle with a simple model of an apple. The performance of the algorithm is investigated using two datasets. The first dataset consists of images recorded in full automatic mode of the camera and under various lighting conditions. Although the algorithm detects correctly more than 85% of the apples visible in the images, direct illumination and color saturation cause a large number of false positive detections. The second dataset consists of images that were manually underexposed and recorded under mostly diffusive light (close to sunset). For such images the correct detection rate is close to 95% while the false positive detection rate is less than 5%.