Research on target recognition, segmentation, and accurate positioning methods for obscured apple fruits under dense canopies

In robotic harvesting, recognizing and locating target apple fruits, as well as detecting obstacles in complex environments like dense canopies, poses significant challenges. This paper presents a method for recognizing, segmenting, and accurately localizing obscured fruits. We employ instance segmentation to extract edge contours of fruits and branches, leveraging two-dimensional features for precise localization under occlusion and correcting picking point errors. Experimental results indicate that at a 10% occlusion ratio, the model achieves an average precision of 96.8%, an F1 Score for edge detection of 0.91, and a 3D reconstruction error of 3.1mm. Even at 50% occlusion, average precision remains at 88.3%, with an F1 Score of 0.80 and a reconstruction error of 7.6mm. These findings demonstrate that while occlusion significantly impacts 3D reconstruction, the algorithm retains high accuracy, fulfilling the precision requirements for fruit harvesting.