【作者】 刘俊；

【导师】 卢安舸；

【作者基本信息】 湘潭大学 ， 机械， 2024， 硕士

【摘要】 当前,莲蓬采摘正面临日渐突显的劳动力短缺,“请人难”问题已成为限制莲子产业可持续发展的阻碍。莲蓬自动采摘是莲子产业的必然发展方向,准确识别莲蓬则是实现该技术的前提。然而,莲蓬生长于非结构环境,其个体大小、生长高度多变,在机器人视野中呈现典型多尺度特征。且存在背景复杂、近色干扰、遮挡频发、光照变化的问题,自动识别难度极大,容易产生误检和漏检。为此,本文提出一种多尺度莲蓬识别方法,建立了莲蓬专用识别模型,在莲蓬采摘机器人上完成了模型部署,并开展了采摘试验验证。论文主要内容如下:(1)莲蓬识别数据集制作与生长性状参数测量。实地采集了非结构环境中的莲蓬图像,制作了多尺度莲蓬识别专用数据集,为深度学习模型建立提供了条件。开展了莲蓬生长性状参数测量试验,结果表明莲蓬生长高度分布范围较宽,外形尺寸不一,且具有随机性,由此造成了莲蓬在图像中呈现多尺度特征。(2)多尺度莲蓬识别模型的建立。基于YOLOv5深度学习算法,建立了一种轻量化的多尺度莲蓬识别专用模型MLP-YOLOv5。改进措施包括:优化了多尺度检测层和锚框参数,引入了C3-TR模块和SA注意力机制,采用了GSConv和VoVGSCSP模块构建轻量化颈部,使用了SIoU损失函数。由此提高了模型对于多尺度莲蓬的检测精度,降低了模型的复杂度。(3)模型性能评价与试验。基于测试集对MLP-YOLOv5模型进行了性能测试与对比。在检测精度方面,模型的准确率P、召回率R和平均精度mAP@0.5值分别为93.7%、90.8%和94.9%,较基线模型分别提高了1.7%、3.9%和3%;此外,模型的权重大小、参数量和计算量分别降低了25.7%、30.0%和14.6%。值得一提的是,模型对小目标的识别准确率P和召回率R分别提高了5.5%和7.4%,有效降低了漏检和误检。与其它主流一阶段算法相比,模型在检测精度、计算规模及检测速度等方面具有优势。(4)基于前期研制的莲蓬采摘机器人开展了实验室环境下的采摘试验。首先,将MLP-YOLOv5模型部署于机器人的Jetson Xavier NX控制器中,并对机器人视觉系统的深度误差进行了检测。通过定位试验校验了采摘机器人对于所识别目标的定位精度。据此,开展了莲蓬自动采摘试验,莲蓬采摘机器人的采摘成功率为93.5%,单个莲蓬的平均采摘时间为11.73 s。上述结果验证了所提出的模型用于莲蓬采摘的实用性和有效性。更多还原

【Abstract】 At present,lotus pod picking is facing an increasingly prominent labor shortage,and the problem of "hard to hire" has become an obstacle to limiting the sustainable development of the lotus seed industry.Automatic picking of lotus pods is the inevitable development direction of the lotus seed industry,and accurate identification of lotus pods is the premise of realizing this technology.However,the lotus pod grows in a non-structural environment,and the lotus pod’s size and growth height are highly variable,showing typical multi-scale characteristics in the harvesting robot’s visual field.In addition,there are problems such as complex background,near-color interference,frequent occlusion,and lighting changes,which make automatic identification extremely difficult and easily produce false and missed detections.To solve these problems,a multi-scale lotus pod identification method is proposed in this dissertation.The model is tested using the built test set.Afterward,the model is deployed on the lotus pod harvesting robot,and the picking test is carried out.The main contents of the paper are as follows:(1)Preparation of the lotus pod identification data set and measurement of the lotus pod growth character parameters.The images of lotus pods in an unstructured planting environment were first collected.Then,an identification data set of multiscale lotus pods was built,which provided conditions for establishing the deep learning model.For another,parameter measurements of lotus pod growth traits were conducted.The results show that lotus pods have the characteristics of a wide distribution range of growth height and variable shapes and sizes among individuals,which results in the multi-scale distribution phenomenon of lotus pods in the captured image.(2)Establishment of the multi-scale lotus pod identification model.Based on the YOLOv5 deep learning algorithm,this dissertation established a lightweight multi-scale lotus pod identification model named MLP-YOLOv5.Several improvement measures were adopted in the model,including optimizing the multiscale detection layer and anchor box parameters,introducing the C3-TR module and SA attention mechanism,using GSConv and VoVGSCSP module to build a lightweight neck,and introducing SIoU loss function.As a result,the detection accuracy of the model for the multi-scale lotus pod was improved,and the complexity of the model was reduced.(3)Performance evaluation of the model.The performance of the MLPYOLOv5 model was tested on the test set.Regarding detection accuracy,the precision(P),recall(R),and mAP@0.5 achieved by the model were 93.7%,90.8%,and 94.9%,respectively,which were 1.7%,3.9%,and 3% higher than the baseline model,respectively.In addition,the model’s weight size,parameters,and GFLOPs were reduced by 25.7%,30.0%,and 14.6%,respectively.It is worth mentioning that the model’s P and R for small objects achieved by the model were increased by 5.5% and7.4%,respectively.The model effectively reduced the phenomenon of false and missed detections in multi-scale object detection.Compared with other mainstream one-stage deep learning algorithms,the model has advantages in detection accuracy,GFLOPs,and detection speed.(4)The laboratory environment picking test was conducted based on the lotus pod harvesting robot developed in the previous stage.Firstly,the MLP-YOLOv5 model was deployed in the Jetson Xavier NX controller of the robot.Then,the depth error of the robot’s depth vision system was verified.Subsequently,the localization accuracy of the harvesting robot to the identified object was verified by the localization test.Based on these,the automatic lotus pod-picking test was carried out.The success rate of the lotus pod-picking robot was 93.5%,and the average picking time of a single lotus pod was 11.73 s.The above results verify the practicability and effectiveness of the proposed model for lotus pod harvesting.更多还原