Estimating crop yields in advance of harvest is an important task in planning crop production operations and marketing. Current yield predictions in apple orchards are based on historical yields in previous years or crop density estimated by sampling one to three limbs per tree. These schemes often produce inaccurate size and yield estimation by failing to account for fruit-to-fruit variation within a tree and tree-to-tree variation within a block. In recent years, there have been various studies on mature and immature fruit detection and size estimation algorithms using machine vision techniques. Many of those studies have focused on recognition of non-occluded fruit using color, texture, or shape. However, partial occlusion of fruit by leaves, branches, and overlapping fruit presents a major challenge in many fruit detection algorithms since the occlusion affects the geometric features of fruit in images.

In this study, an innovative method to detect partially occluded apples on trees will be developed using a CHOI’s Circle Estimation (‘CHOICE’) algorithm. The CHOICE, originally developed to identify non-occluded immature citrus, uses depth images to detect spherical objects in three-dimensional (3D) space. In the algorithm, the 3D geometric features of an object surface can be calculated using a gradient vector field of a depth map. Analysing different patterns of the gradient vector field of depth in partially occluded fruit, this study will focus on improving the CHOICE algorithm to identify apple fruit with occlusion and accurately estimate the size of apples in trees at various distances from the sensor to tree canopy. A computer vision system will be developed using a Kinect V2 sensor to utilize depth, color, and near-infrared images. Various fruit occlusion scenarios will be simulated in indoor conditions to develop a machine vision algorithms. The results of this study will be analyzed in terms of false positives and false negatives generated by the machine vision algorithm. Also, a root-means-square-error (RMSE) between actual sizes and estimated sizes of apples will be provided. Using the proposed machine vision system, the in-field spatial variability of apple size and count can be provided immediately. Using this information, apple growers can implement site-specific management by modifying their operations according to the specific crop yield at different locations in their orchards