【作者】 冯健；

【导师】 马俊燕； 正端；

【作者基本信息】 广西大学 ， 机械工程（专业学位）， 2022， 硕士

【摘要】 我国是农业生产大国,每年生产的水果总量在世界名列前茅。柑橘作为我国产量第二大的水果产业,产量在逐年递增的同时面临劳动力的流失,采用高新技术助力农业生产具有重要意义。本文针对静态图像获取柑橘产量时可能产生的重复计数问题,采用无人机航拍视频方式识别果实并预测产量,视频拍摄时只需无人机搭载的摄像头按不重复路线扫描一遍即可获取果园果实产量信息,且可用于丘陵地区复杂地形的场景。为了实现从视频中获取柑橘产量信息,本文的主要研究内容如下:(1)采集果园中不同拍摄角度、不同距离的柑橘图像800张用于柑橘识别算法的训练和测试,以6:1:1的比例划分为训练集、验证集和测试集,并采用增加噪声和Mosaic对训练图像进行数据增强以提高训练数据量。同时,通过无人机平行树行飞行拍摄的方式不重复的采集果园视频,并人工统计视频中的果实数量用于产量估算方法的测试数据。(2)为使产量估算方法可以实时运行,本文对YOLOv5目标检测算法进行轻量化改进。采用轻量计算单元Ghost module和通道注意力机制ECA-Net重构主干网络中的瓶颈层,并将主干网络中的标准卷积替换为深度可分离卷积。改进后的模型参数量为1.83×10~6,仅有原模型的四分之一,大大降低了模型复杂度。(3)为使模型在参数量降低后保证检测精度。采用Soft-NMS算法解决密集检测时的漏检问题,提高检测召回率。训练阶段,采用遗传算法在多次迭代训练中获得最优超参数组合。并对多种损失函数进行实验测试,选用CIo U作为训练的损失函数,以提高模型的训练效果。使用柑橘数据集对充分训练后的模型进行测试,取得了96.6%的m AP,模型运行速度可达131Fps,可见在保证一定识别准确率的前提下,明显提升了检测速度,增大了模型的可用性。(4)采用DeepSORT目标跟踪算法将跟踪器与果实目标一一匹配,为每个果实打上独立的ID,从而获取视频中的果实数量得到柑橘产量信息。为了更好的提取柑橘外观特征,使用Res Net-18作为外观特征提取网络。并在柑橘识别算法中加入边界框阈值以规避背景果实对计数的干扰。将该方法应用于多段不同角度拍摄、包含不同果实数量的柑橘园视频中,获得的准确率保持在82%以上,最高可达92.9%,结果表明本文提出的柑橘产量估算方法在多种情况下仍可保持较高的准确率,具有一定的实用价值,可为柑橘生产和物流规划提供有力的技术支持。更多还原

【Abstract】 China is a large agricultural country,and the total amount of fruit produced every year ranks among the best in the world.As the second largest fruit industry in China,citrus is facing the loss of labor force while its output is increasing year by year.It is of great significance to use high and new technology to help agricultural production.Aiming at the problem of repeated counting that may occur when still images obtain citrus yield,this paper uses UAV aerial video to identify the fruit and predict the yield.During video shooting,only the camera carried by UAV can scan once according to the non repeated route to obtain the orchard fruit yield information,and can be used in the scene of complex terrain in hilly areas.In order to obtain citrus yield information from video,the main research contents of this paper are as follows:(1)Collect citrus images from different shooting angles and distances in the orchard,use 800 for the training and testing of Citrus recognition algorithm,divide them into training set,verification set and test set in the proportion of 6:1:1,and enhance the data of training images by adding noise and mosaic to improve the amount of training data.At the same time,the orchard video is not repeatedly collected by means of UAV parallel tree flight shooting,and the number of fruits in the video is manually counted for the test data of yield estimation method.(2)In order to make the yield estimation method run in real time,this paper improves the YOLOv5 target detection algorithm.The lightweight computing unit ghost module and channel attention mechanism ECA-Net were used to reconstruct the bottleneck layer in the backbone,and the standard convolution in the backbone was replaced by deep separable convolution.The parameter of the improved model is 1.83×10~6,only a quarter of the original model,which greatly reduces the complexity of the model.(3)In order to ensure the detection accuracy of the model after reducing the parameters.Soft-NMS algorithm was used to solve the problem of missing detection in intensive detection and improve the detection recall rate.In the training stage,genetic algorithm was used to obtain the optimal combination of super parameters in multiple iterative training.A variety of loss functions are tested experimentally,and CIo U was selected as the training loss function to improve the training effect of the model.Using citrus data set to test the fully trained model,96.6%map is obtained,and the running speed of the model can reach 131fps.It can be seen that on the premise of ensuring a certain recognition accuracy,the detection speed is significantly improved and the usability of the model is increased.(4)The DeepSORT target tracking algorithm was used to match the tracker with the fruit target one by one,and each fruit is marked with an independent ID,so as to obtain the number of fruits in the video and obtain the citrus yield information.In order to better extract the appearance features of citrus,Res Net-18 was used as the appearance feature extraction network.The boundary box threshold was added to the citrus recognition algorithm to avoid the interference of background fruit on the count.The method is applied to many citrus orchard videos taken from different angles and containing different fruit quantities,and the accuracy rate remains above 82%,up to 92.9%.The results show that the citrus yield estimation method proposed in this paper can still maintain high accuracy in many cases,has certain practical value,and can provide strong technical support for citrus production and logistics planning.更多还原