Integration of Instance Segmentation and Three-Dimensional Point Cloud Analysis for Pineapple Pose Estimation

Fruit pose estimation technology is a critical component for enabling intelligent agricultural harvesting, holding significant application value for improving picking efficiency and reducing labor costs. To address the issues of high complexity in existing instance segmentation models and insufficient real-time performance in 3D pose estimation, this paper proposes a lightweight instance segmentation model, FLUG-YOLO, and an efficient method for pineapple pose estimation. Based on the YOLOv8-seg architecture, computational redundancy is reduced by replacing the backbone network with FasterNet. The feature representation capability is enhanced by embedding the Large-Separable-Kernel Attention (LSKA) mechanism within the feature pyramid. The neck convolutional layers are reconstructed using GhostConv, and a novel C2f_UIB structure is designed by incorporating the Universal Inverted Bottleneck (UIB) module. These modifications significantly reduce the number of parameters while improving multi-scale feature fusion performance.Furthermore, by integrating 3D point cloud analysis, a Principal Component Analysis (PCA)-based method for rapid pose estimation is proposed. This method calculates the pineapple's roll and pitch angles by determining the principal axes of the point cloud. Experimental results demonstrate that FLUG-YOLO achieves an Average Precision (AP) of 96.1% (a 2.2% improvement over the baseline model) and an F1-score of 92.6% (a 3.1% improvement) on a custom pineapple dataset, with a model size of only 3.2 MB (a 49.2% reduction). For pose estimation, the average errors in unoccluded scenarios were 3.29° (roll) and 3.52° (pitch), increasing to 6.19° (roll) and 7.88° (pitch) under mild occlusion. The average pose estimation processing time was 301 ms. This approach achieves a synergistic optimization of accuracy, lightweight design, and speed, providing valuable technical support for intelligent pineapple harvesting.