自然果园果实三维点云语义分割模型OA

[Objective]3D point cloud data of fruits in natural orchard scenarios are characterized by uneven distribution,disor-dered structures,and class imbalance.Existing semantic segmentation methods are mostly designed for regular scenes,suffer-ing from insufficient local feature representation and limited capabilities in modelling long-distance semantic relationships in com-plex agricultural environments,making it difficult to achieve high-precision segmentation of fruits from branches and leaves.Therefore,this paper proposed OrchardNet,a semantic segmentation model for 3D point clouds of fruits in natural orchard sce-narios.[Methods]A hierarchical encoder-decoder structure was constructed based on the PointNet++framework.First,point cloud data was preprocessed using methods such as scene segmentation,data augmentation,and spatial sampling.Sec-ond,a Local Feature Aggregation(LFA)module was designed,which normalized neighborhood feature offsets using a Geo-metric Differentiated Scaling(GDS)strategy to enhance the model's capability to represent small-scale fruit structures.Further-more,a Global Feature Mapping(GFM)module was constructed,introducing a Position Encoding Self-Attention(PESA)mechanism to capture long-distance spatial semantic dependencies in the point cloud.Simultaneously,a Weighted Cross-Entro-py Loss function was employed to alleviate the class imbalance problem between the fruit and the background.[Results]Experi-ments were conducted on the PFuji-Size public dataset.The results showed that the OrchardNet model achieved a mean Inter-section over Union(mIoU)of 87.4%,which was 5.9%and 3.3%higher than PointNet++(SSG)and PointNet++(MSG),respectively,and 9.4%and 2.1%higher than Point Cloud Transformer and PointMLP,respectively.The mean Ac-curacy(mAcc)reached 95.5%,which was 6.0%and 3.6%higher than PointNet++(SSG)and PointNet++(MSG),re-spectively.[Conclusion]The OrchardNet mode effectively improved the semantic segmentation accuracy of 3D point clouds of fruits in natural orchards,and provided reliable 3D perception technology support for tasks such as automatic harvesting by agri-cultural robots,fruit growth monitoring,and yield assessment.