【作者】 赵鹏；

【导师】 杨蜀秦；

【作者基本信息】 西北农林科技大学 ， 机械， 2025， 硕士

【摘要】 草莓苗田间杂草繁殖迅速,不仅会争夺营养和光照、升高局部环境温度,还会成为病虫害的中间宿主,加速和扩大病虫害的发生和传播。本研究围绕草莓种苗工厂化田间育苗场景,针对田间杂草精准识别与高效除草需求,设计并实现了一套集视觉检测、自主导航与激光除草于一体的智能机器人。通过理论分析、算法优化与系统集成,主要工作如下:(1)激光除草机器人总体方案确定。基于实地调研与理论分析,对激光除草机器人进行了总体方案设计,主要由调节龙门履带底盘、激光控制系统和视觉检测系统组成。其中可调节龙门履带底盘充分考虑草莓苗田间垄作环境及不平整地形问题,设计并计算了履带底盘结构和主要参数,根据可调节龙门履带底盘总体结构设计和作业不同工况下阻力的理论计算与分析,确定了选用额定电压48V直流无刷电机,并搭配蜗轮蜗杆减速机,为可调节龙门履带底盘供能,确保机器人在狭窄通道和复杂地形中的高机动性与稳定性;激光控制系统,选用风冷激光器搭配高功率振镜,确保了激光控制系统在高效能、精准定位和安全防护方面均能满足激光除草机器人的实际作业需求,为系统整体性能优化提供了有力保障;视觉检测系统,由检测主机和D435i深度相机构成,确保实时运行目标识别算法并提供高质量图像数据。(2)针对草莓田间杂草与草莓苗尺寸较小、颜色相近,滴灌管形状细长,以及包含杂草生长点复合目标定位问题,提出基于改进YOLOv8s-pose的DIN-LW-YOLO模型。该模型在YOLOv8-pose的高分辨率特征图上构建预测头。在预测头和快速空间金字塔模块(SPPF)之前,加入EMA注意力模块,用于捕获像素级的成对关系,显著提升了像素级关注能力和小目标特征保留效果;同时针对细长型滴灌管目标,利用可变形卷积自适应捕获目标的特征信息,用其替换特征融合模块中瓶颈结构的第二个卷积,捕获更丰富的特征信息。目标检测方面,DIN-LW-YOLO模型比YOLO5s-pose、YOLOv6s-pose、YOLOv8s-pose模型精度分别提高3.5%、2.8%、2.9%,其余指标优于其他目标检测模型。点目标检测方面,本研究的DIN-LW-YOLO检测模型的指标均优于其他检测模型,模型参数量也较低。此模型对目标检测与关键点检测的综合性能优于其他模型,能够同时实现对草莓苗、杂草、滴灌管的区域目标和对杂草生长点目标的准确检测。(3)针对草莓田间田垄中垄面、垄沟目标边缘易受背景影响,垄沟与支撑柱细长型目标分割时的边缘断裂问题,提出基于YOLOv10s-seg改进的FR-Nav-YOLO模型。该模型在YOLOv10s-seg基础上引入Re-Calibration FPN重校准特征金字塔网络,通过整合P3、P4、P5多尺度特征与新增高分辨率P6特征图,增强了垄面区域特征提取能力;同时采用基于残差注意力块的特征融合模块,优化支撑柱检测过程,有效提升了细长型垄沟与边缘完整性。实验结果显示,FR-Nav-YOLO在区域目标检测中实现了98.9%的精确率和96.6%的召回率,m AP达到了98.2%,F1分数为97.7%。与其他模型相比,FR-Nav-YOLO不仅在检测性能上更胜一筹,而且在实时性方面也具有较高的优势,能够在复杂场景中快速响应,实现精细分割目标的实时检测。(4)在草莓田间杂草检测方法DIN-LW-YOLO和田间地形检测方法FR-Nav-YOLO研究的基础上,基于Bo T-SORT跟踪算法与D435i深度相机构建激光靶点动态定位系统,结合坐标变换技术解决相机与振镜视野差异问题;提出一种多算法协同导航控制策略,机器人可沿滴灌管及田垄自主行驶,实现复杂地形中的动态避障与路径修正;同时,基于Py Qt6开发的交互式控制软件利用CAN总线与TCP协议实现底盘、激光与视觉系统的闭环控制。实地验证表明,该系统杂草控制率达92.6%,草莓苗损伤率仅为1.2%,为草莓田间自动化激光除草提供了一种有效的方案。更多还原

【Abstract】 Weeds in strawberry seedling fields proliferate rapidly.They not only compete for nutrients and sunlight and raise the local temperature,but also serve as intermediate hosts for pests and diseases,thereby accelerating and expanding their occurrence and spread.This study focuses on the field nursery scenario of industrialized strawberry seedling production and addresses the need for precise weed identification and efficient removal.An intelligent robot system integrating visual detection,autonomous navigation,and laser weeding was designed and implemented.The main contributions of this work,achieved through theoretical analysis,algorithm optimization,and system integration,are as follows:(1)Based on field investigation and theoretical analysis,an overall design scheme for the laser weeding robot was proposed.The system mainly consists of an adjustable gantry crawler chassis,a laser control system,and a visual detection system.The adjustable gantry crawler chassis is specifically designed to adapt to the ridged layout and uneven terrain commonly found in strawberry fields.Its structure and key parameters were designed and calculated to ensure optimal performance.Through analysis of structural layout and theoretical resistance under different working conditions,a 48V brushless DC motor combined with a worm gear reducer was selected to power the chassis,ensuring high maneuverability and stability in narrow paths and complex terrains.The laser control system employs an air-cooled laser source paired with a high-power galvanometer scanner,providing high efficiency,precise targeting,and reliable safety features to meet the practical demands of laser-based weeding operations,while also enhancing overall system performance.The visual detection system,consisting of a processing unit and a D435i depth camera,ensures real-time execution of target recognition algorithms and supplies high-quality image data for accurate environmental perception.(2)To address the challenges of detecting small-sized weeds and strawberry seedlings with similar colors,as well as the elongated shape of drip irrigation pipes and the complex localization of weed growth points,this study proposes the DIN-LW-YOLO model,an improved version of YOLOv8s-pose.The model integrates an Enhanced Mean Attention(EMA)module before the prediction head and the Spatial Pyramid Pooling-Fast(SPPF)module to capture pixel-level relationships,significantly improving small target feature retention.Additionally,deformable convolution was incorporated into the feature fusion module,replacing the second convolution in the bottleneck structure,enhancing feature extraction for elongated drip irrigation pipes.Compared with YOLO5s-pose,YOLOv6s-pose,and YOLOv8s-pose,DIN-LW-YOLO improves detection accuracy by 3.5%,2.8%,and 2.9%,respectively,and outperforms other models in all key performance indicators.For point target detection,DIN-LW-YOLO demonstrates superior accuracy and lower parameter complexity,making it highly effective in identifying strawberry seedlings,weeds,and drip irrigation pipes,as well as accurately localizing weed growth points.(3)For challenges in distinguishing ridge surfaces and furrow edges in strawberry fields,as well as the edge fragmentation issues in segmenting elongated furrows and support columns,this study proposes the FR-Nav-YOLO model,an improvement based on YOLOv10s-seg.This model incorporates the Re-Calibration Feature Pyramid Network(Re-Calibration FPN)to integrate P3,P4,and P5 multi-scale features with a newly introduced high-resolution P6feature map,enhancing ridge surface feature extraction.Additionally,a residual attention block-based feature fusion module was introduced to optimize support column detection and improve the continuity of elongated furrow boundaries.Experimental results show that FR-Nav-YOLO achieves a precision of 98.9%,a recall of 96.6%,a mean Average Precision(m AP)of 98.2%,and an F1-score of 97.7%.Compared with other models,FR-Nav-YOLO offers superior detection performance and real-time efficiency,enabling fast response in complex environments and precise segmentation of targets.(4)Building upon the DIN-LW-YOLO weed detection model and the FR-Nav-YOLO terrain detection model,a dynamic laser targeting system was developed using the Bo T-SORT tracking algorithm and the D435i depth camera,resolving vision differences between the camera and the galvanometer through coordinate transformation techniques.A multi-algorithm collaborative navigation control strategy was introduced,enabling the robot to autonomously navigate along drip irrigation pipes and field ridges,dynamically avoiding obstacles and correcting its path in complex terrains.Additionally,an interactive control software was developed based on Py Qt6,using CAN bus and TCP protocols to achieve closed-loop control of the chassis,laser,and vision systems.Field tests demonstrated a weed control rate of 92.6%and a strawberry seedling damage rate of only 1.2%,providing an effective solution for automated laser weeding in strawberry fields.更多还原