An Automatic Segmentation and Recognition Method of Apple Tree Point clouds in the Real Scene Based on the Fusion of Color and 3D Feature

Abstract. It is a challenging task to achieve accurately fruit recognition in 3D space for the robotic fruit picking. This paper proposed an automatic segmentation and recognition method for the robotic task of apple picking from the point cloud data. First, an HSI based region growing segmentation method was proposed to segment the point cloud data of apple tree. Next, an improved 3D descriptor with the fusion of color features and 3D geometry features was computed from each segmented point clouds. And then, a support vector machine optimized by genetic algorithm (GA-SVM) classifier was constructed to recognize apples, branches and leaves in the scene. We applied quantitative results and lateral comparison to demonstrate and evaluate the performance of the proposed method. The result of recognition accuracy was 92.30%, 88.03%, 80.34 for apples, branches, leaves, respectively and the lateral comparison results showed the proposed method had better robustness and generalization performance.