Research on fruit 3D point cloud segmentation based on improved YOLOv8

With the development and progress of science and technology in our country, the continuous improvement of people's living standards, the planting area of fruits is constantly expanding, and the fruit yield is also constantly increasing, resulting in a shortage of labor resources and bringing huge challenges to fruit picking; At the same time, due to the impact of different geographical environments, climates, and other natural environments, as well as the mutual occlusion of fruits and leaves, the positioning and picking of fruits are affected, requiring a large amount of manpower to complete the picking operation. This article uses the improved YOLOv8 algorithm for instance segmentation of fruits to improve the accuracy and efficiency of instance segmentation. By extracting mask information, corresponding depth information is obtained, and the point cloud data corresponding to apples is obtained, providing assurance for the accuracy of subsequent localization. Firstly, in the backbone network, the deformable convolutional DCNv2 is replaced with the C2f module to increase attention to small targets; By adding an improved ECA attention mechanism EMCA before the Neck layer, the model efficiency is improved. Finally, by proposing WPDIoU instead of CIoU to improve the loss, the convergence speed and segmentation accuracy of the model are improved. After experimental comparison between the original YOLOv8 model and the improved YOLOv8 model proposed in this paper, the results show that the improved YOLOv8 model has higher instance segmentation accuracy, an increase of 4.2 percentage points, and has a certain improvement in fruit instance segmentation accuracy.