摘要
针对田间路径语义分割算法在复杂场景中识别准确率低、时滞大,现有的导航线提取算法在三七种植场景中适用性差的问题,本文以改进的轻量化Deeplab-MV3模型识别垄径区域,并提出一种结合分段外接框质心生成导航定位点和多段三次B样条曲线的导航中线提取方法,实现对三七垄径导航线的准确高效提取.在三七田间数据集上训练模型并通过梯度加权类激活映射法分析Deeplab-MV3的各部分模块改进的有效性,测试集上模型垄间识别的平均交并比为74.80%,像素准确率为94.53%,参数量为3.086×10~6,模型推理速度为55.44帧/s,满足部署到移动设备的识别准确性和模型轻量化要求.开展三七垄间导航线提取实验,结果表明本文算法的平均像素偏差为5.32像素,误差占比的平均偏差为0.91%,偏航角平均偏差为1.27°,相较于使用Canny算子边缘提取的中点的导航线拟合法的导航精度有所提高.实验结果可为复杂道路田间环境的导航线提取以及农机导航设备提供研究基础.
In response to the low accuracy and significant latency of field path semantic segmentation algorithms in complex scenarios,as well as the limited applicability of existing navigation line extraction algorithms in the Panax notoginseng plantation scenes,this paper introduces an improved lightweight Deeplab-MV3 model for identifying the monopoly path region.It proposes a navigation line extraction method that combines the segmented bounding box centroid to generate navigation points and multi-segment cubic B-spline curves to extract the navigation lines of Panax notoginseng ridges with accuracy and efficiency.The model was trained on the Panax notoginseng field data sets and the improving effectiveness of each module in Deeplab-MV3 was analyzed by gradient weighted class activation mapping method.The mean intersection over the union of the model inter-monopoly recognition on the test sets was 74.80%,the pixel accuracy was 94.53%,the number of the parameters was 3.086 ×10~6,and the model inference speed was 55.44 frames per second,satisfying the recognition accuracy and model lightweight requirements deployed to mobile devices.The experiment on the navigation line extraction of Panax notoginseng ridges was carried out.The results showed that the average pixel deviation of the proposed algorithm was 5.32 pixels,the average deviation of the error percentage was 0.91%,and the average deviation of the yaw angle was 1.27°,which improved the navigation accuracy compared with the navigation line fitting method using the midpoint of the Canny operator edge extraction.The experimental results can serve as a research foundation for navigation line extraction in complex field road environments and for agricultural machinery navigation equipment.
作者
陈红
张兆国
解开婷
王成琳
王法安
CHEN Hong;ZHANG Zhaoguo;XIE Kaiting;WANG Chengin;WANG Faan(Faculty of Modern Agricultural Engineering,Kunming University of Science and Technology,Kunming 650093,China;Mechanization Engineering Research Center of Chinese Medicinal Materials in Yunnan Univercities,Kunming 650093,China;Faculty of Mechanical and Electrical Engineering,Kunming University of Science and Technology,Kunming 650500,China)
出处
《昆明理工大学学报（自然科学版）》
北大核心
2023年第5期95-106,共12页
Journal of Kunming University of Science and Technology(Natural Science)
基金
国家重点研发计划项目(2022YFD2002004)
院士(专家)工作站项目(202105AF150030)。
关键词
三七
语义分割
Deeplab-MV3
垄间导航线
梯度加权类激活映射
Panax notoginseng
semantic segmentation
Deeplab-MV3
field navigation lines
gradient-weighted class activation mapping
