摘要
[目的]针对目前农业机器人在非结构化枣园中传统机器视觉算法难以处理光影变化、夜间难以识别等问题导致的导航道路分割不准确,实时性及复杂模型难以部署等问题,本研究提出基于YOLOv5s-seg改进网络模型视觉导航路径识别方法——YOLOv5s-orchard模型。[方法]模型采用FasterNet代替YOLOv5s-seg的原有主干网络,减少了模型的参数量和内存占用量;在特征提取层及neck部分引入SE注意力机制,提高了模型的特征提取能力;将C3替换成C3Ghost轻量级网络以进一步减少模型的参数量和内存占用量。[结果]改进后的网络mAP达到91.59%,FLOP下降到11.30 G,模型大小为2.91 MB。对分割后的图像采用最小二乘法拟合道路离散点提取导航线,算法平均距离偏差为1.99 cm,平均距离误差比例为0.68%。[结论]YOLOv5s-orchard网络具有较高的精度,导航线提取误差较小,能够满足枣园道路分割的需求。
[Objectives]In view of the current agricultural robots in the unstructured jujube garden,the traditional machine vision algorithm has difficulty dealing with light and shadow changes,has difficulty identifying objects at night,leading to inaccurate navigation road segmentation,real-time and has difficulty deploying complex models.In this study,we proposed a modified network model visual navigation path recognition method based on YOLOv5s-seg-YOLOv5s-orchard model.[Methods]The model used FasterNet instead of the original backbone network of YOLOv5s-seg,which reduced the number of parameters and memory usage of the model;introduced the SE attention mechanism in the feature extraction layer and neck part to improve the feature extraction ability of the model;replaced C3 with C3Ghost lightweight network to further reduce the number of parameters and memory usage of the model.[Results]The results showed that the improved network mAP reached 91.59%,FLOP decreased to 11.30 G and a model size of 2.91 MB.The navigation lines were fitted by least squares to the segmented images.The mean distance deviation of the algorithm was 1.99 cm,and the mean distance error ratio was 0.68%.[Conclusions]The results indicated that the YOLOv5s-orchard network had high accuracy.The navigation line extraction error was relatively small.The YOLOv5s-orchard could meet the needs of jujube orchard road segmentation.
作者
卢心缘
赵华民
任锐
苏萌
苏立阳
张淑娟
LU Xinyuan;ZHAO Huamin;REN Rui;SU Meng;SU Liyang;ZHANG Shujuan(College of Agricultural Engineering/Machinery Key Technology and Equipment Key Laboratory of Shanxi Province,Shanxi Agricultural University,Taigu 030801,China)
出处
《南京农业大学学报》
CAS
CSCD
北大核心
2024年第6期1241-1250,共10页
Journal of Nanjing Agricultural University
基金
山西省研究生实践创新项目(2023SJ100)
山西省自然科学基金青年项目(202103021223145)。
