基于RT-DETR视觉定位和机械臂路径规划的马铃薯分拣系统OA
Potato Sorting System Based on RT-DETR Visual Positioning and Manipulator Path Planning
为了解决马铃薯收获环节仍依赖大量人工,且现有的分段收获机无法有效完成分选和清理工作等问题,设计了基于深度学习的目标检测与机械臂路径规划的马铃薯分拣系统.该系统由视觉定位和机械臂路径规划两部分组成.视觉定位由RT-DETR算法识别马铃薯、土块和石块,完成对应目标的定位.通过相机标定与手眼标定实现世界坐标系、相机坐标系、图像坐标系及像素坐标系之间的转换,并基于传送带高度获取目标空间位置坐标.运动学分析基于Denavit-Hartenberg参数建立机械臂连杆坐标系,通过齐次变换矩阵实现正运动学建模,并采用几何法简化逆运动学求解,推导出关节角解析表达式.路径规划选用三次多项式插值法,通过设定起始终止角度、速度边界条件,生成平滑轨迹,确保各关节加速度变化平缓.最终,马铃薯分拣系统基于PyTorch框架部署RT-DETR轻量化检测模型,模型数据集上准确率达到93.3%,召回率达到93.0%,在堆叠、覆土、强光条件下分别实现97.5%、90.3%、95%的识别准确率.集成OpenCV与Socket通信协议,系统定位误差≤1.342 mm.结合渐进式分拣策略与三次多项式轨迹规划,取得92.9%分拣成功率与1.3%误检率.马铃薯分拣系统能够在不同条件下对马铃薯与杂物精准识别,并协同机械臂实现马铃薯实验台分拣.
With the advancement of the potato staple food strategy in China,the cultivation area of potatoes has been expanding.However,existing segmented potato harvesters still face challenges in sorting and cleaning efficiency,with a high reliance on manual labor.Particularly,during the sorting process,it is difficult to effectively distinguish potatoes from soil clods and stones of similar shapes and sizes based on physical characteristics,leading to continued dependence on manual sorting at conveyor belts.To address this issue,this study designed a potato sorting system based on the deep learning target detection and robotic arm path planning,aim-ing to reduce the labor intensity and improve the sorting efficiency through automation technology.This would promote the wide-spread application of segmented potato harvesters and support the implementation of the potato staple food strategy.The potato sort-ing system consisted of two main components,i.e.,visual positioning and robotic arm path planning.The visual positioning module employed the RT-DETR(Real-Time DEtection TRansformer)algorithm to identify potatoes,soil clods,and stones,and to locate the targets.The camera calibration and hand-eye calibration were utilized to achieve the transformation between the world coordinate system,camera coordinate system,image coordinate system,and pixel coordinate system,while the spatial coordinates of the tar-gets were obtained based on the conveyor belt height.The robotic arm path planning module established the robotic arm link coordi-nate system using Denavit-Hartenberg(D-H)parameters,realized forward kinematic modeling through homogeneous transformation matrices,and simplified inverse kinematic solutions using a geometric approach to derive analytical expressions for joint angles.Path planning was performed using cubic polynomial interpolation,genarating smooth trajectories by setting initial and final angles and velocity boundary conditions,ensuring smooth acceleration changes in each joint for precise robotic arm movement.Experimen-tal results demonstrated that the RT-DETR lightweight detection model deployed on the PyTorch framework achieved an accuracy of 93.3%and a recall rate of 93%on the dataset.Under conditions of stacking,soil covering,and strong light,the system achieved recognition accuracies of 97.5%,90.3%,and 95%,respectively.By integrating the OpenCV and the Socket communication proto-col,the system's positioning error was controlled to≤1.342 mm.Combined with a progressive sorting strategy and cubic polynomial trajectory planning,the system achieved a sorting success rate of 92.9%and a false detection rate of 1.3%.The experiments validat-ed the system's robustness under various environmental conditions,effectively addressing challenges such as target overlap,morpho-logical variations,and environmental interference.Additionally,the system's average sorting cycle was 8.7 seconds per item,dem-onstrating its practical application value.The successful application of this system provided a feasible solution for the automated sort-ing of segmented potato harvesters,demonstrating significant practical value.Future work may focus on optimizing algorithms and hardware configurations to enhance the system's real-time performance and stability for more complex field operations.
孔祥;吴英思;刘飞;玄德正;张安斌
内蒙古农业大学机电工程学院,呼和浩特市 010018内蒙古农业大学机电工程学院,呼和浩特市 010018内蒙古农业大学机电工程学院,呼和浩特市 010018内蒙古农业大学机电工程学院,呼和浩特市 010018内蒙古农业大学机电工程学院,呼和浩特市 010018
农业科技
马铃薯除杂深度学习目标检测路径规划机械臂分拣
Potato impurity removalDeep learningTarget detectionPath planningMechanical arm sorting
《内蒙古农业大学学报(自然科学版)》 2026 (1)
66-76,11
内蒙古自治区自然科学基金项目(2021MS05067)
评论