Classification of Pickable and Unpickable Strawberries under Farm Conditions

In this paper, we address the problem of deformed point clouds of small targets under farm condition. Strawberry point cloud data were extracted through image detection and corresponding coordinate transformation. The deformed point cloud caused by adhesion of adjacent objects make the localization inaccurate. Therefore, multiple features in three spatial dimensions were constructed to represent the shape and distribution of the strawberry point data. Some fundamental features, such as eccentricity and least moment of inertia, were utilized, while some new features, such as 3D shape matrix and ratio of data amount in certain space, were proposed. Furthermore, classifiers that can identify unpickable strawberry data and other defined classes were trained and evaluated on constructed features and different label sets. The evaluation results show that the classifier could effectively identify unpickable cases. This work proposes a new problem and provides possible solutions that could optimize the performance of machine vision system of the strawberry harvesting robot.