Design of an Intelligent Pear Bagging End-Effector Based on Yolov8 and SGBM Algorithm

Abstract

Bagging is a crucial step in the full cycle management of pear cultivation, with increasing labor costs annually, driving research on intelligent bagging equipment. We describe an intelligent bagging end-effector for pears, which employs the Yolov8 algorithm for fruitlets detection and the Semi-Global Block Matching (SGBM) algorithm to acquire three-dimensional spatial information of the targets. To address the computational limitations of embedded devices in agricultural intelligent equipment, we improve the YOLOv8 model by replacing its neck component with the Asymptotic Feature Pyramid Network (AFPN) and incorporating Context Guided (CG) blocks into the C2f module. These improvements significantly reduce the model’s parameters and size while enhancing detection accuracy. To provide accurate three-dimensional information to the end-effector, we use the SGBM algorithm to obtain depth information of the targets. We constrain the matching process based on the model’s detected candidate bounding boxes and binocular prior knowledge, which improves the matching accuracy of the model while reducing the computational load. Finally, we deploy the model and matching algorithm on a Jetson Nano, which serves as the controller for the bagging end-effector based on the vacuum negative pressure principle. Experimental results show that the improved model achieves a precision of 96%, an increase of 1.4% over YOLOv8, with the size reduced from 6.2M to 5.0M. The depth information accuracy obtained by the algorithm is within 3cm. The improved model, combined with SGBM, runs at 11.5 fps on the Jetson Nano embedded device, meeting real-time performance requirements.

Keywords

• Embedded devices
• lightweight model
• depth information