基于语义分割的山地果茶园道路识别技术研究

Semantic segmentation based road recognition technology of hilly fruit and tea garden

针对果、茶园规模不断扩张并逐渐向智能农业机械化发展的趋势以及常用道路语义分割数据集缺少果、茶园道路场景等问题,将语义分割技术应用到部分果、茶园道路中,以实现对果、茶园道路的像素级分割。以道路、人和车为分类对象,建立果、茶园道路场景图像数据集(包括6032张图像),将数据集按照9∶1比例随机划分为训练集(5429张图像)和测试集(603张图像)。以PSPNet(pyramid scene parsing network,金字塔场景解析网络)分割模型为基础进行优化,构建MS-PSPNet语义分割模型;训练结果显示,MS-PSPNet模型的MIoU(mean intersection over union,平均交并比)为83.41%,FPS(frames per second,每秒传输帧数)为22.31。将MS-PSPNet模型应用在果、茶园不同道路条件和光照强度下进行现场试验,并进行准确度评估,结果显示,MS-PSPNet模型类别MPA(mean pixel accuracy),像素准确率均超过92%,MIoU在除非硬化道路条件情况均超过91%,表明MS-PSPNet模型在果、茶园道路识别中具有较好的有效性和适用性。

Aiming at the trend of continuous expansion of fruit and tea gardens,the gradual development of intelligent agricultural mechanization and the lack of fruit and tea garden road scenes in common road semantic segmentation data sets,the semantic segmentation technology was used to some fruit and tea garden roads in Meizhou City,Guangdong Province to realize pixel-level segmentation of roads in fruit and tea gardens.Roads,people,and cars were used as classification objects to establish a scene image data set including 6032 images of fruit and tea garden road.The data set was randomly divided into a training set including 5429 images and a test set including 603 images according to a 9∶1 ratio.The MS-PSPNet semantic segmentation model was established based on the PSPNet(pyramid scene parsing network)segmentation model for optimization.The results of training showed that MS-PSPNet model mean intersection over union(mean intersection over union,MIoU)was 83.41%.The number of frames per second(frames per second,FPS)was 22.31.The MS-PSPNet model was applied to fruit and tea gardens under different road conditions and light intensity to conduct field tests and evaluate the accuracy.The results showed that the category pixel accuracy(mean pixel accuracy,MPA)of MS-PSPNet model exceeded 92%.MioU exceeded 91%in all cases of non-hardened road conditions.It is indicated that the MS-PSPNet model has good validity and applicability in road recognition of fruit and tea gardens.